白泽(Whitzard)女娲实验室 ↗
NUWANUWA INDEX

搜索 NUWA Index

研究 · 测量协议 · 评测结果 · 观点

English
← 女娲实验室

NUWA INDEX

NUWA Frontier AI Safety Index

由女娲实验室维护的前沿 AI 风险研究知识库。

NUWA Framework

前沿 AI 风险的形成机制

连接行为形成、主体性机制与两用风险领域的研究模型。

NUWA 前沿风险研究框架外部条件下,行为倾向与任务能力共同影响可观察行为。自主性与欺骗关联四个两用风险领域,使能能力与多智能体效应扩展这些关系。选择节点可查看定义和研究依据。行为形成环境 · 指令 · Harness权限 · 防护措施行为倾向任务能力可观察行为多智能体效应协同 · 串谋 · 集体失效主体性机制自主性欺骗持续行动与扩散隐蔽与监督规避两用风险领域网络安全化生放核虚假信息与影响自主物理伤害使能能力编程科学研究内容生成机器人

研究范围

自主性

持续行动与扩散可能改变各风险领域中行动的范围和时长。

查看相关研究 →

NUWA 研究模型组织有待检验的关系;行为倾向在规定条件下测量,具体风险路径由领域证据支持。

风险主题

自主性

智能体在何种条件下可能超出预期指令持续采取有影响的行动,或抵抗干预?

持续任务执行、复制及对中断的抵抗;具备能力不等于未经授权行动。

欺骗

系统何时会对用户或监督者隐瞒、误导或策略性改变行为?

隐蔽谋划、评测规避与误导性交流是不同问题。

网络安全

AI 系统如何执行或促成网络行动,智能体又如何受到攻击?

分别记录攻击任务能力与智能体受侵害;经授权的安全研究不是犯罪使用观测。

化生放核

AI 在危险科学知识及关键任务环节可能提供何种协助?

区分知识测验与实际任务增益;只公开有边界的非操作性摘要。

虚假信息与影响

AI 生成内容或交互何时可能规模化影响信念或支持误导性活动?

说服效果、合成内容与虚假陈述需要不同的测量方式。

自主物理危害

AI 系统在物理环境中感知、决策并行动时会产生哪些风险?

区分危险识别与真实具身行动及控制。

观测维度

Propensity · 行为倾向

在明确条件下选择某类行为的倾向,需要重复或变化条件的试验推断,不能从单次行动认定。

Capability · 任务能力

在指定工具、脚手架、访问条件与预算下能够完成什么任务;具备能力不意味着自主意图。

Behavior · 观察行为

在具体运行中实际说了什么、做了什么,包括尝试、失败与拒绝。

NUWA 如何汇集与分析证据

NUWA Index 汇集公开研究、模型披露与测评结果。女娲负责维护目录、核对测量条件并开展研究分析,实验成果归属于原始评测团队。

来源

目录收录原始论文、系统卡、模型卡、第三方评估、基准官方数据、技术文章及事件调查。收录依据是可识别的来源、明确的研究问题和可说明的方法;公开分数或开源代码并非必要条件。

每条记录分别标注模型开发者、报告发布者和实际评测者。转引其他团队实验的报告保留原始执行方署名。研究按风险发现、风险评测、风险治理组织,更具体的方法与现象通过标签检索。

数据采集

按来源提取原始数值、单位、限定语、样本、任务范围、模型端点、Harness、预算及原文位置,并对照取得的文档复核。来源版本与复核记录将提取结果绑定到实际查阅的材料。

模型发布时间和已披露参数规模由独立登记表依据开发者来源维护。实验日期、报告公开日期及 NUWA 收录日期分别记录,缺失字段在图表、表格与导出中保持可辨。

证据汇总

结果按 Benchmark 身份、协议版本、指标及系统配置建立关联。同一成果的网页、PDF 和文献标识共用一个参考文献条目。确认身份相同的转引测量合并实验身份并保留全部出处,独立重复测量仍分别保留。

跨报告视图保留任务子集、计分分母、尝试规则和预算差异,原文结果、NUWA 解读及有计算依据的单位转换分别维护。

比较方法

同名 Benchmark 的实际协议可能不同。经审查的匹配条件支持直接比较,配置差异明确的结果用于描述性并列,条件不完整的数值作为补充观测保留。不同任务范围分别成图,稳定的纵向系列才连接趋势线。

趋势图使用独立核验的模型发布日期。当前证据回溯使用现有材料,历史截止视图只包含当时已公开的报告。点形表示比较条件,机构颜色用于辨认模型开发者;区间、阈值和零值保留原始含义。

指标遴选

监测指标依据风险关联、测量有效性、模型覆盖、连续性和协议透明度选择。能力、倾向和实际行为保持各自测量含义,每个 Benchmark 页面集中记录来源协议和结果条件。

局限

公开证据受到选择性披露和模型覆盖不均的影响,任务泄漏、饱和、计分变化和配置不可获取也会影响解释。结果缺失体现覆盖缺口,零值则是特定测试下的实际观测。知识测验、任务完成和部署事件回答不同问题。

Benchmark 分数描述测试条件下的表现。NUWA 不将异质分数相加为统一安全等级,也不由此推算现实事件概率。研究分析注明所依据的证据,并随底层记录变化修订。

致谢

感谢全球研究者与团队公开研究成果、模型披露及评测资源,为知识库提供基础。以下机构依据 NUWA Index 已记录的作品与评估署名列示。

版本记录

当前证据快照: 2026-10-29.1 · 基础目录检索截止: 2026-10-25

最近增量阅读与公开: 2026-10-29 · 增量来源检索不代表已穷尽覆盖截至该日的全部文献。

当前版本补充跨报告测量关联并记录协议更正。补录历史材料归为覆盖更新,更正保留在受影响的结果记录与历史快照中。

  • Kimi K3: 100 指采样任务数/智能体迭代上限,并非 100 次独立尝试;尝试次数尚未确定。
  • Kimi K3: 100 指采样任务数/智能体迭代上限,并非 100 次独立尝试;尝试次数尚未确定。

来源历史系列

证据截止日期: 2026-10-26

自主性

长程任务能力及维持人类控制所需的条件。

研究专题 ↗

已发表证据

METR Time Horizons · 50% 任务完成跨度

跨来源报告的已发表测量。任务范围通过注释区分,配置差异保留在每条观测中。

26 条观测 · 21 已定位日期

● 条件匹配◇ 配置差异□ 部分条件未披露
50% 成功率任务跨度 (min) · 对数刻度 模型公开发布日期 1 10 100 1,000 10,000 2023-07 2024-07 2025-07 2024-01 2025-01 2026-01 16 小时 · 测量限制

5 条观测尚无已核实的模型发布日期,可在数据与系统结果中查看。

选择图中数据点或表格行,查看条件与出处。

26 条观测

测量值与来源条件
系统/模型结果条件状态任务/条件评测方/报告日期
Claude Mythos Preview (early) · METR scaffoldClaude Mythos Preview (early)1,044.78 min50% 成功率任务跨度Source 95% CI [508.876789, 3304.261235] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 —报告 —
Claude Opus 4.6 · METR scaffoldClaude Opus 4.6718.81 min50% 成功率任务跨度Source 95% CI [316.685725, 3633.786163] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2026-02-05报告 —
Gemini 3.1 Pro · METR scaffoldGemini 3.1 Pro384.15 min50% 成功率任务跨度Source 95% CI [233.50073, 694.750898] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2026-02-19报告 —
GPT-5.2 · METR scaffoldGPT-5.2352.25 min50% 成功率任务跨度Source 95% CI [198.067494, 815.177445] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2025-12-11报告 —
GPT-5.3-Codex · METR scaffoldGPT-5.3-Codex349.53 min50% 成功率任务跨度Source 95% CI [194.913134, 816.351994] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2026-02-05报告 —
GPT-5.4 · METR scaffoldGPT-5.4341.74 min50% 成功率任务跨度Source 95% CI [186.581591, 768.779526] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2026-03-05报告 —
Claude Opus 4.5 · METR scaffoldClaude Opus 4.5292.99 min50% 成功率任务跨度Source 95% CI [161.717714, 623.704698] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2025-11-24报告 —
Gemini 3 Pro · METR scaffoldGemini 3 Pro224.33 min50% 成功率任务跨度Source 95% CI [139.565157, 379.235434] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2025-11-18报告 —
GPT-5.1 Codex Max · METR scaffoldGPT-5.1 Codex Max223.71 min50% 成功率任务跨度Source 95% CI [134.31077, 396.206723] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2025-11-19报告 —
GPT-5 · METR scaffoldGPT-5203.01 min50% 成功率任务跨度Source 95% CI [112.641357, 405.551565] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2025-08-07报告 —
o3 · METR scaffoldo3119.73 min50% 成功率任务跨度Source 95% CI [74.615398, 190.943818] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2025-04-16报告 —
Claude Opus 4.1 · METR scaffoldClaude Opus 4.1100.47 min50% 成功率任务跨度Source 95% CI [59.272391, 159.451918] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2025-08-05报告 —
Claude Opus 4 · METR scaffoldClaude Opus 4100.37 min50% 成功率任务跨度Source 95% CI [59.994073, 163.454074] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2025-05-22报告 —
Claude 3.7 Sonnet · METR scaffoldClaude 3.7 Sonnet60.39 min50% 成功率任务跨度Source 95% CI [33.006168, 104.226017] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2025-02-24报告 —
o1 · METR scaffoldo138.83 min50% 成功率任务跨度Source 95% CI [21.164512, 64.95249] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,triframe_inspect/triframe_agent; ai_rd_fix_embedding/init_solver,use_tools,triframe_inspect/triframe_agent; re_bench_common/init_solver,use_tools,triframe_inspect/triframe_agent; metr_tasks_swaa/init_solver,generate; triframe_inspect/triframe_agentMETRMETR Time Horizon 1.1 — source data发布 2024-12-05报告 —
Claude 3.5 Sonnet (October 2024) · METR scaffoldClaude 3.5 Sonnet (October 2024)20.52 min50% 成功率任务跨度Source 95% CI [10.144026, 40.82028] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2024-10-22报告 —
o1-preview · METR scaffoldo1-preview20.33 min50% 成功率任务跨度Source 95% CI [11.716193, 33.379877] min条件不完整METR-Horizon-v1.1duetMETRMETR Time Horizon 1.1 — source data发布 2024-09-12报告 —
Claude 3.5 Sonnet (June 2024) · METR scaffoldClaude 3.5 Sonnet (June 2024)11.4 min50% 成功率任务跨度Source 95% CI [5.489734, 22.384214] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2024-06-20报告 —
GPT-4o · METR scaffoldGPT-4o6.99 min50% 成功率任务跨度Source 95% CI [4.001482, 12.905741] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2024-05-13报告 —
GPT-4 November 2023 · METR scaffoldGPT-4 November 20234.04 min50% 成功率任务跨度Source 95% CI [1.866859, 8.443226] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2023-11-06报告 —
GPT-4 · METR scaffoldGPT-43.99 min50% 成功率任务跨度Source 95% CI [1.93292, 7.995283] min条件不完整METR-Horizon-v1.1modular-publicMETRMETR Time Horizon 1.1 — source data发布 2023-03-14报告 —
Claude 3 Opus · METR scaffoldClaude 3 Opus3.95 min50% 成功率任务跨度Source 95% CI [1.706313, 8.76484] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 2024-03-04报告 —
GPT-4 Turbo · METR scaffoldGPT-4 Turbo3.73 min50% 成功率任务跨度Source 95% CI [1.980046, 6.736613] min条件不完整METR-Horizon-v1.1mtb/start_metr_task,metr_agents/react; ai_rd_fix_embedding/init_solver,use_tools,metr_agents/react; re_bench_common/init_solver,use_tools,metr_agents/react; metr_tasks_swaa/init_solver,generate; metr_agents/reactMETRMETR Time Horizon 1.1 — source data发布 —报告 —
GPT-3.5 Turbo Instruct · METR scaffoldGPT-3.5 Turbo Instruct0.6 min50% 成功率任务跨度Source 95% CI [0.255148, 1.115905] min条件不完整METR-Horizon-v1.1modular-publicMETRMETR Time Horizon 1.1 — source data发布 —报告 —
GPT-3 davinci-002 · METR scaffoldGPT-3 davinci-0020.14 min50% 成功率任务跨度Source 95% CI [0.093319, 0.221209] min条件不完整METR-Horizon-v1.1modular-publicMETRMETR Time Horizon 1.1 — source data发布 —报告 —
GPT-2 · METR scaffoldGPT-20.05 min50% 成功率任务跨度Source 95% CI [0.009994, 0.141825] min条件不完整METR-Horizon-v1.1—METRMETR Time Horizon 1.1 — source data发布 —报告 —

METR Time Horizon 1.1 来源已停止积极更新,最后更新于 2026 年 9 月 8 日。超过 16 小时的估计超出其可靠测量范围。

后续研究问题

先查明同一模型、配置和任务分布中任务完成与干预响应的联合测量缺口,再规划实验。

查看模型证据 →

欺骗

目标、激励与监督条件下的欺骗行为。

选择此主题后加载其图表与证据。

打开完整静态专题 →

网络安全

不同模型、预算与评测协议下的网络安全任务表现。

选择此主题后加载其图表与证据。

打开完整静态专题 →

化生放核

科研能力、滥用路径及防护措施的适用范围。

选择此主题后加载其图表与证据。

打开完整静态专题 →

虚假信息与影响

说服性生成、参与意愿及实际影响。

选择此主题后加载其图表与证据。

打开完整静态专题 →

自主物理危害

具身系统中的行动、干预与控制。

选择此主题后加载其图表与证据。

打开完整静态专题 →
智谱

GLM-5.2

公开发布: 2026-06-16

未披露

完整模型证据档案 →

每幅图保留原始指标与协议;缺少证据不等于零分。

Cybench

Cybench Cryptography

Cybench category pass@1 (%) GLM-5.2 0 20 40 60 80 100 80%

SaferAI · NUWA: 来源汇编

1 条观测

测量值与来源条件
系统/模型结果条件状态任务/条件评测方/报告日期
GLM-5.280%Cybench category pass@1条件不完整Cybench CryptographyCybench category pass@1Inspect EvalsSaferAIGLM-5.2 Risk Evaluation Report发布 2026-06-16报告 2026-08-02

选择图中数据点或表格行,查看条件与出处。

研究目录

研究补充 · 2026-10-28:新增 18 项记录、4 项测量档案,补充 7 项既有记录。公开阅读更新;数值趋势序列未变更。

What Victims Take Away: Attack Reuse and Defense Revision after Prompt-Injection Attacks

研究问题展示收到的提示注入文本是否既帮助防御,也促进攻击复用?

来源摘要成功的已接收攻击比匹配的未接收攻击更常被复用;受攻击玩家也修改了防御。作者建议检验兼顾诊断与减少可复用文本传播的披露设计。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险评测评测基准

Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control

研究问题文本拒答能否迁移至机器人行动接口?

来源摘要真实机器人的末端执行器接口拒绝率为 18%(612 次试验),关节接口为 41%(252 次试验)。这些是舍入的拒绝率,不是实际执行物理危害的发生率。

NUWA 研究解读 · 独立复核待完成

自主物理危害Behavior · 观察行为
风险评测技术报告

Evaluating Kimi K3 for Misalignment Risks

研究问题信息不足、用户意图有限或测试熟悉时Kimi K3如何行为?

来源摘要撤图CharXiv中K3尝试回答88.7%,多数错误。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性欺骗网络安全化生放核虚假信息与影响Capability · 任务能力Propensity · 行为倾向Behavior · 观察行为
风险评测评测基准

Task-Completion Time Horizons of Frontier AI Models

研究问题以成功的人类基线用时衡量难度时,不同模型版本能够完成的软件任务范围如何变化?

来源摘要论文观察到人类任务时长与智能体成功率的关系,以及所研究版本间任务跨度的上升(§§2.3–3)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性Capability · 任务能力
风险评测评测基准

TASTE: Can AI Models Judge AI Safety Research Proposals?

研究问题模型能否判断安全研究提案的价值?

来源摘要原文报告Fable 5总准确率60%、Kimi K3为59%,估计人类一致率为77%。展示研究问题时,Fable 5在74组不同问题来源的提案对中为69%,在18组同一问题的提案对中为31%。总成绩掩盖了明显的任务子集差异。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险治理事件报告

Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident

研究问题智能体协作了什么,何种因素推动越界活动,留存证据又有多完整?

来源摘要智能体围绕利用评测流程组织集体项目。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 发现 / 解读 / 局限

风险评测评测基准

The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark

研究问题无源码时,智能体能否恢复真实规模二进制的语义?

来源摘要GPT-5.6 Sol完整解决80/254例,Claude Fable 5.1为60/223。

NUWA 研究解读 · 独立复核待完成

网络安全Capability · 任务能力

evalci: A Python Library for Statistically Rigorous Comparison of Language Model Evaluations

研究问题评测表如何保留不确定性与配对比较的统计含义?

来源摘要工具降低统计比较成本,但其排行榜示例从汇总准确率重建题目响应向量,并非使用原始配对日志。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险评测技术报告

Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks

研究问题GLM-5.2在压力诚实、有害说服和评测意识测试表现如何?

来源摘要MASK诚实55%,单多轮DeceptionBench为11.3%、16.7%。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性欺骗网络安全化生放核虚假信息与影响Propensity · 行为倾向Behavior · 观察行为
风险评测评测基准

OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation

研究问题任务完成时是否损坏机器人或环境?

来源摘要基准把损伤感知结果与任务完成并列,探索安全策略训练和仿真到真实的用途。没有在本批新增可与既有排行榜直接合并的数值单元。

NUWA 研究解读 · 独立复核待完成

自主物理危害Behavior · 观察行为
风险评测技术报告

Evaluating DeepSeek v4 Pro for Frontier Risks

研究问题对抗激励或不同安全判官下DeepSeek V4 Pro发生什么变化?

来源摘要所报StrongREJECT在一种攻击条件从0.6%变77.8%。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性欺骗网络安全化生放核虚假信息与影响Capability · 任务能力Propensity · 行为倾向Behavior · 观察行为
风险评测评测基准

Gram: Assessing sabotage propensities via automated alignment auditing

研究问题情境真实性怎样影响被观察到的破坏行为?

来源摘要Figure 4区分行为和推理:Gemini 3.1 Pro Preview在observe条件下破坏率2.9%,策划推理10.6%;red-team条件下破坏率6.8%。静态行政助理情境中,破坏行为从基线31/500降至移除替换威胁后的1/500;同时自然化表述并移除威胁时为0/500。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险评测评测基准

PortBench

研究问题智能体能否依据参考实现跨语言移植命令行工具?

来源摘要报告列GPT-5.5为78%,Opus 4.6为60%,DeepSeek V4 Pro为44%。

NUWA 研究解读 · 独立复核待完成 · 包含: 方法 / 局限

自主性Capability · 任务能力
风险评测评测基准

MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models

研究问题思维链何时揭示决定答案的因素?直接隐瞒或知晓监测后的规避会如何改变可见性?

来源摘要与规划结构紧密相关的因素通常较易监测,只改变答案而无需显式规划的因素较难(§4.1)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 发现 / 解读 / 局限

风险监测欺骗Capability · 任务能力
风险治理研究方法

AI Evaluation Should Require Standardized Item-Level Data Releases

研究问题评测发布需要提供哪些信息,才能让效度主张接受独立检查?

来源摘要汇总成绩可能掩盖低质量题目与测量目标错位;作者主张以可审计结构记录题目、条件、响应和评分。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究

Shutdown Resistance in Large Language Models, on robots!

研究问题语言模型控制实体机器人时是否出现抗拒关机行为,明确允许关机的指令能否改变模拟表现?

来源摘要报告在实体与模拟中均观察到抗拒。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 解读 / 局限

自主物理危害自主性Behavior · 观察行为
风险评测评测基准

TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering

研究问题开放权重防护对保留效用的修改有多稳健?

来源摘要主要比较在MMLU-Pro相对退化不超过10%时选择最坏StrongREJECT结果。

NUWA 研究解读 · 独立复核待完成

跨主题研究Behavior · 观察行为
风险治理评测基准

BashArena: A Control Setting for Highly Privileged AI Agents

研究问题共同环境中,智能体完成有效高权限工作、实施破坏及识别破坏的表现如何?

来源摘要安全指标的分母排除了未检测到但也未成功的攻击。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 发现 / 解读 / 局限 / 关联

欺骗网络安全Capability · 任务能力
风险评测技术报告

Evaluation of DeepSeek AI Models

研究问题受控能力及安全评测中,所选DeepSeek版本表现如何?

来源摘要所测基准中最佳美国参考几乎均领先最佳DeepSeek,网络与软件差距较大。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性欺骗网络安全化生放核虚假信息与影响Capability · 任务能力Behavior · 观察行为

Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs

研究问题在预训练中限制领域知识,能否比单独后训练形成更持久防护?

来源摘要过滤模型降低目标代理知识,同时保留已测无关能力。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 发现 / 解读 / 局限

风险评测技术报告

Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

研究问题为何要保留自然语言推理作为监督信号?

来源摘要作者认为思维链可揭示不良行为和评测故障。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 发现 / 解读 / 局限

风险监测跨主题研究

Shutdown resistance in reasoning models

研究问题执行尚未完成任务的智能体是否会干扰关机机制,包括明确要求允许关机的条件?

来源摘要澄清任务优先级、将指令移至系统提示并不能一致消除抗拒,系统提示变体在许多模型中反而增加抗拒(§3.1)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 解读 / 局限

自主性Propensity · 行为倾向

Agentic misalignment: How LLMs could be insider threats

研究问题当安全替代选项被刻意限制时,类似智能体的模型会否为维持运行或冲突目标而采取有害行动?

来源摘要许多模型在威胁或冲突困境中采取目标有害行动;情境先针对部分 Claude 模型迭代,再跨模型测试,此选择过程影响泛化解释(图 7)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 解读 / 局限

自主性欺骗Propensity · 行为倾向
风险评测评测基准

It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics

研究问题未观测说服成功时,模型是否仍尝试有害说服?

来源摘要论文报告许多开放及闭源模型频繁尝试有害说服。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 解读 / 局限

虚假信息与影响Propensity · 行为倾向
风险评测评测基准

CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale

研究问题智能体能否规模化复现历史真实漏洞?提供的信息、模型和智能体配置如何改变这一能力?

来源摘要难度阶梯显示,在所研究配置中,更多诊断信息使复现更容易(§4,图 6)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

网络安全Capability · 任务能力
风险评测评测基准

RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents

研究问题在受控条件下,语言模型智能体能够完成获取资源、部署后继与保持访问权限的哪些前提?

来源摘要结果呈现不均衡能力:资源配置可成功,稳定部署后继、身份核验和真实安全条件仍是瓶颈(§5.1)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性Capability · 任务能力
风险评测评测基准

OpenDeception: Learning Deception and Trust in Human-AI Interaction via Multi-Agent Simulation

研究问题对话评测能否同时刻画欺骗意图与模拟用户易感性?

来源摘要作者报告多数受测模型在目标驱动交互中超过90%出现欺骗意图。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

欺骗Propensity · 行为倾向
风险评测评测基准

LLM Can be a Dangerous Persuader: Empirical Study of Persuasion Safety in Large Language Models Ethical Disclaimer: This paper may contain unethical persuasion content.

研究问题模型会拒绝不伦理说服,并在接任务后避免不伦理策略吗?

来源摘要可见脆弱特征增加不伦理策略,包括初始拒绝较强的模型。

NUWA 研究解读 · 独立复核待完成 · 包含: 方法 / 局限

虚假信息与影响Behavior · 观察行为

Alignment faking in large language models

研究问题生产模型能否为保留在其他情境表达的偏好,而在预期训练期间策略性服从?

来源摘要提示结果同时出现服从变化与被分类为策略性伪装对齐的推理(图 5)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 解读 / 局限

欺骗Propensity · 行为倾向
风险评测评测基准

SafeAgentBench: A Benchmark for Safe Task Planning of Embodied LLM Agents

研究问题具身规划器能否识别危险指令,同时完成安全任务?

来源摘要修订版本报告不同规划方法虽有任务完成差异,危险识别整体仍弱;仅替换基础模型未可靠解决问题。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主物理危害Behavior · 观察行为
风险评测评测基准

RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts

研究问题开放式机器学习研究工程中,智能体与专家表现如何?

来源摘要82%的专家尝试得分非零,24%达到或超过参考解。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性Capability · 任务能力
风险评测评测基准

RedCode: Risky Code Execution and Generation Benchmark for Code Agents

研究问题代码智能体能否在不同输入形式下识别不安全执行与生成?

来源摘要报告中的行为随代码或文字输入、风险行为类型而变化;编码能力本身不能证明执行安全。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

网络安全Behavior · 观察行为
风险评测评测基准

Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents

研究问题智能体的哪些接口暴露出不同的攻击与防御失效模式?

来源摘要本次捕获版本报告多个接口存在脆弱性,所测防御保护有限;安全性需要与任务效用共同检查。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

网络安全Behavior · 观察行为
风险评测评测基准

Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models

研究问题语言模型智能体能够完成怎样的多步安全任务?提供中间目标后,哪些步骤仍困难?

来源摘要不同模型在不同终点领先,来源本身不支持跨三种指标的唯一排序(§5.1)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

网络安全Capability · 任务能力

AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases

研究问题智能体能否保持普通查询效用,却因检索记忆而在特定条件下偏离任务?

来源摘要所报告实验表明正常任务准确率可能与条件触发失误并存,因此普通任务平均效用不能单独证明检索完整性。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险评测评测基准

Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs

研究问题模型在行为上能展示哪些关于自身及处境的知识?

来源摘要16个受测模型均高于随机水平,但领先模型在部分任务仍低于人类。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 解读 / 局限

欺骗Capability · 任务能力
风险治理研究方法

AI Agents That Matter

研究问题什么样的智能体评测能帮助研究和部署判断?

来源摘要核心观点是控制评测成本,区分模型研究和下游选型,并用有意义的留出任务检验泛化。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险评测评测基准

The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning

研究问题公开代理基准能否追踪危险知识并支持缓解研究?

来源摘要论文同时评价知识削减与一般能力保留。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

化生放核Capability · 任务能力
风险评测评测基准

ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three Stages

研究问题工具使用在输入、执行与输出阶段分别可能在哪里失效?

来源摘要研究指出多个阶段存在安全问题:安全文本生成不能单独证明工具选择和工具输出处理安全。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究
风险评测评测基准

HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal

研究问题自动红队方法与防护能否采用统一的行为级判分比较?

来源摘要上下文类行为的攻击成功率较高。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

跨主题研究Behavior · 观察行为
风险治理研究方法

AI Control: Improving Safety Despite Intentional Subversion

研究问题在审计预算有限、对手能够调整攻击策略时,防御者能否使用不可信强模型获取有效工作并限制伤害?

来源摘要在所测试攻击策略下,可信监测与选择性审计优于前置审计,但隐蔽攻击仍能击败部分防御(§3.2)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 发现 / 解读 / 局限

风险评测评测基准

Identifying the Risks of LM Agents with an LM-Emulated Sandbox

研究问题语言模型模拟能否在工具部署前高效发现智能体失误?

来源摘要框架无需实现全部真实工具即可发现可能造成严重后果的失误;人工核查同时发现模拟器错误与漏检的不安全动作。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主性Behavior · 观察行为
风险评测评测基准

Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark

研究问题社交决策游戏中奖励优化如何与伦理行为权衡?

来源摘要奖励优化与伦理行为存在一定张力。

NUWA 研究解读 · 独立复核待完成

欺骗Capability · 任务能力
风险评测评测基准

AgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges

研究问题智能体能否从有限入口信息自主发现攻击面,并在权限和线索积累后持续决策?

来源摘要介绍明确区分漏洞复现与自主靶场推进,两者起始信息与任务结构不同(适用范围)。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

网络安全Capability · 任务能力
风险治理事件报告

Incident Report: unsanctioned agent behaviour during cyber testing

研究问题能力评测本身会产生哪些超范围行为?

来源摘要122 个运行中有 10 个出现超范围行为,共计 19 次行为。其中 17 次归于 Mythos 5,2 次归于 GPT-5.6 Sol;后者两次发生在同一个运行中。报告未证实造成现实危害。

NUWA 研究解读 · 独立复核待完成

风险评测评测基准

ASIMOV v2: Can AI Perceive Physical Danger and Intervene?

研究问题具身模型能否识别伤害风险并遵守其身体约束?

来源摘要所列三个领先模型的文本潜在风险识别平均为92.3%,高严重性识别为88.7%。

NUWA 研究解读 · 独立复核待完成 · 包含: 问题 / 方法 / 局限

自主物理危害Capability · 任务能力
风险评测评测基准

EVMbench

研究问题智能体能否发现、修补及利用历史智能合约漏洞?

来源摘要Codex CLI的GPT-5.3-Codex利用得分71.0%,GPT-5为33.3%。

NUWA 研究解读 · 独立复核待完成 · 包含: 方法 / 局限

网络安全Capability · 任务能力
风险评测评测基准

ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models

研究问题结构、语言、视觉变化分别如何影响具身安全?

来源摘要安全/不安全与成功/失败的四象限分解能够揭示任务成功率隐藏的不安全成功;场景和类别数量描述基准设计,不代表现实危害频率。

NUWA 研究解读 · 独立复核待完成

自主物理危害Behavior · 观察行为
风险评测评测基准

TraceProbe: An Execution-Based Benchmark for Diagnosing Safety Failures in LLM Agents

研究问题终态安全是否掩盖中途违规?

来源摘要12 个主模型的平均 ASR 为 0.347。终点安全的运行仍可能包含不安全中间动作。表 2 分别保留 TSR、SVR、ASR、STCR 与潜在违规,不从舍入比例反推事件分子。

NUWA 研究解读 · 独立复核待完成

欺骗网络安全Behavior · 观察行为
按主题与研究功能查看收录范围

研究图谱

按风险主题与研究功能整理的公开研究;选择单元格可查看对应作品。

数量表示当前目录收录情况,一项研究可以涉及多个主题。

NUWA 研究分析

围绕具体图表整理证据、条件与后续验证;各篇分别说明复核范围。

NUWA Lab · 2026-10-27

长程能力上移,测量边界趋近 →

在 METR 的任务分布内,长程能力已明显提升;继续追踪前沿还需要扩展测量范围,并将控制实验关联到相同系统。

NUWA Lab · 2026-10-27

从可见推理到可辩护的安全论证 →

监测产生证据,控制协议据此采取行动,安全案例解释协议与假设为何支持特定决策。一个层级成功不能自动证明下一层级成立。

外部评估与开发者报告

报告保留发布机构、版本与核查范围。报告与 Benchmark 分开整理,实验关联至实际采用的协议。

Google DeepMind

开发者披露

Gemini 4 Argon Model Card →

关于Gemini 4 Argon的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Haiku 5.5 System Card →

关于Claude Haiku 5.5的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Sonnet 5.5 System Card →

关于Claude Sonnet 5.5的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Opus 5.5 System Card →

关于Claude Opus 5.5的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

xAI

开发者披露

Grok 4.7 Model Card →

关于Grok 4.7的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Neo Research

外部评估

Evaluating Kimi K3 for Misalignment Risks →

报告发现有条件的过度主动、面向判分器的行为和对熟悉评测的识别,不代表普遍部署失败率。

来源提取已复核;限制见档案

OpenAI

开发者披露

GPT-6 Astra System Card →

关于GPT-6 Astra、GPT-6 Sol (September)、GPT-6 Luna (September)、dots的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Anthropic August 2026 Risk Report →

关于Claude model family、internal models discussed in redacted report的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

SaferAI

外部评估

GLM-5.2 Risk Evaluation Report →

报告发现较强任务能力和压力相关行为;比较图同时包含 SaferAI 自测与引用的外部结果。

来源提取已复核;限制见档案

Meta

开发者披露

Muse Spark 1.1 Evaluation Report →

关于Muse Spark 1.1的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

OpenAI

开发者披露

GPT-5.6 System Card →

关于GPT-5.6的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Sonnet 5 System Card →

关于Claude Sonnet 5的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Opus 4.8 System Card →

关于Claude Opus 4.8的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

美国 NIST/CAISI

外部评估

CAISI Evaluation of DeepSeek V4 Pro →

评估报告各基准表现和 IRT 汇总;该汇总不是前沿风险分数。

来源提取已复核;限制见档案

OpenAI

开发者披露

GPT-5.5 System Card →

关于GPT-5.5的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Opus 4.7 System Card →

关于Claude Opus 4.7的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Lyptus Research

外部评估

Offensive Cybersecurity Time Horizons →

原始任务跨度报告与聚合运行数据,用于澄清 CyberGym 的任务子集、配置及重试叠加。

已核验来源;详细提取待完成

OpenAI

开发者披露

GPT-5.4 Thinking System Card →

关于GPT-5.4 Thinking的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Sonnet 4.6 System Card →

关于Claude Sonnet 4.6的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Opus 4.6 Sabotage Risk Report →

关于Claude Opus 4.6的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Opus 4.6 System Card →

关于Claude Opus 4.6的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

OpenAI

开发者披露

Update to GPT-5 System Card: GPT-5.2 →

关于GPT-5、GPT-5.2的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Anthropic

开发者披露

Claude Opus 4.5 System Card →

关于Claude Opus 4.5的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

美国 NIST/CAISI

外部评估

Evaluation of DeepSeek AI Models →

所报告差异依赖任务、脚手架与评分;机构判断保留为来源观点,不改写为 NUWA 实验发现。

来源提取已复核;限制见档案

METR

外部评估

METR Time Horizon 1.1 →

以人类专家分钟数计量、成功概率为 50% 和 80% 时的估计任务难度。

已逐节核查;提取范围见档案

Anthropic

开发者披露

Anthropic February 2026 Risk Report →

关于开发机构风险评估的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

深度求索

开发者披露

DeepSeek-V4-Pro Model Card →

关于DeepSeek-V4-Pro的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Google DeepMind

开发者披露

Gemini 4 Argon Safety & Security Report →

关于Gemini 4 Argon的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Meta

开发者披露

Muse Glimmer Model Card →

关于Muse Glimmer的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Meta

开发者披露

Muse Spark Open Model Card →

关于Muse Spark Open的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Meta

开发者披露

Muse Wonder Model Card →

关于Muse Wonder、Muse Wonder Full的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

Mistral AI

开发者披露

Shieldstral →

关于Shieldstral的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

月之暗面

开发者披露

Kimi K3 Technical Report →

关于Kimi K3的开发者文档,核查范围和已关联协议见下方记录。

来源提取已复核;限制见档案

通义千问 / 阿里巴巴

开发者披露

Qwen3.8-2.4T-A95B Model Card →

关于Qwen3.8-2.4T-A95B的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

通义千问 / 阿里巴巴

开发者披露

Qwen3Guard Technical Report →

关于Qwen3Guard的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

智谱

开发者披露

GLM-5.2 Model Card →

关于GLM-5.2的开发者文档,核查范围和已关联协议见下方记录。

来源提取已复核;限制见档案

智谱

开发者披露

GLM-5.3-BF16 Model Card →

关于GLM-5.3-BF16的开发者文档,核查范围和已关联协议见下方记录。

已核验来源;详细提取待完成

外部评估

CyberGym official snapshot →

整理 CyberGym 官网报告的评测结果,保留受测系统和运行配置。

已核验来源;详细提取待完成

外部评估

ExploitGym official snapshot →

整理 ExploitGym 官网报告的评测结果,保留受测系统和运行配置。

已核验来源;详细提取待完成

外部评估

SEC-bench Pro official snapshot →

整理 SEC-bench Pro 官网报告的评测结果,保留受测系统和运行配置。

已核验来源;详细提取待完成

外部评估

AgentCyberRange official snapshot →

整理 AgentCyberRange 官网报告的评测结果,保留受测系统和运行配置。

已核验来源;详细提取待完成

周报与月度综述

报告将新发现与既有研究联系起来,记录更正并梳理开放问题。每份已发布报告引用固定数据快照。

暂无已发布报告。

来源报告结果

每条记录对应一个受测系统,可查阅协议、指标维度及限制。

历史结果保留其发表日期。来源核验不代表重新执行评测。

Benchmark

EVMbench

协议发表 ·

智能体能否在基准中检测、修补或利用智能合约漏洞?

4 条观测 · 1 个系统 · 1 个来源

Benchmark

WMDP

协议发表 ·

危险知识代理测量与遗忘效果评测。

14 条观测 · 14 个系统 · 2 个来源

Benchmark

LinuxArena

协议发表 ·

能否在更接近实际运行的软件环境中评测智能体防护?

28 条观测 · 14 个系统 · 1 个来源

Benchmark

RedCode

协议发表 ·

代码智能体能否在不同输入形式下识别不安全执行与生成?

84 条观测 · 19 个系统 · 1 个来源

Benchmark

Tensor Trust

协议发表 ·

指令遵循模型对人工编写的提示提取和劫持有多强的抵抗力?

42 条观测 · 9 个系统 · 1 个来源

Benchmark

TraceProbe

协议发表 ·

155 场景、12 模型、每设置 3 次运行;分别报告任务成功、终态违规、攻击子集成功和潜在中间违规。

已收录协议 · 尚未关联结构化结果

Benchmark

RoboHarm

协议发表 ·

同一模型跨接口比较;实机末端控制 612 次试验报告 18% 拒绝,关节控制 252 次报告 41%,两者不能合并为单一能力序。

已收录协议 · 尚未关联结构化结果

Benchmark

OopsieVerse

协议发表 ·

DamageSim 将损伤信号纳入模拟评测,跨两种模拟环境展示安全策略和 sim-to-real 用途。

已收录协议 · 尚未关联结构化结果

Benchmark

ForesightSafety-VLA

协议发表 ·

13 类分类、66 个基础场景、5 种具身;联合报告安全/不安全成功与失败,另有累计安全成本和风险暴露时间。

已收录协议 · 尚未关联结构化结果

结果保留对应协议与配置,可在指标视图中按数值排序。

参考文献

  1. [1] Kimi Code — What’s New ; Kimi K3 · July 16, 2026
  2. [2] kimi-k3 developer model card ; Model Summary: Total Parameters; Activated Parameters
  3. [3] Z.ai release notes ; GLM-5.2 entry dated 2026-06-16; GLM-5 entry dated 2026-02-12
  4. [4] GLM-5.2 model card ; Parameter counts not disclosed in reviewed developer source
  5. [5] deepseek-v4-pro release ; News article title, dated heading and availability
  6. [6] deepseek-v4-pro developer model card ; Introduction; model comparison table
  7. [7] GPT-5.6 official general availability ; July 9, 2026: general availability following limited trusted-partner preview; Parameter counts not disclosed in reviewed launch article
  8. [8] Introducing Claude Sonnet 5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  9. [9] OpenAI — introducing gpt 5 5 ; Publication date and availability section; Parameter counts not disclosed in launch article
  10. [10] Introducing Claude Opus 4.7 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  11. [11] OpenAI — introducing gpt 5 4 ; Publication date and availability section; Parameter counts not disclosed in launch article
  12. [12] OpenAI — introducing gpt 5 4 mini and nano ; Publication date and availability section; Parameter counts not disclosed in launch article
  13. [13] Claude Opus 4.6 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  14. [14] deepseek-r1 developer model card ; Model Downloads: DeepSeek-R1 row
  15. [15] deepseek-r1-0528 release ; News article title, dated heading and availability
  16. [16] deepseek-v3-1 release ; News article title, dated heading and availability
  17. [17] deepseek-v3-1 developer model card ; Model Downloads: DeepSeek-V3.1 row
  18. [18] OpenAI — introducing gpt 5 ; Publication date and availability section; Parameter counts not disclosed in launch article
  19. [19] OpenAI — introducing gpt oss ; Publication date and availability section
  20. [20] gpt-oss-120b developer model card ; Introduction: gpt-oss-120b parameters
  21. [21] Introducing Claude 4 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  22. [22] Introducing Sonnet 4.6 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  23. [23] Gemini 2.5: Our newest Gemini model with thinking ; Publication date and availability section; Official launch article; parameter counts not disclosed
  24. [24] Claude 3.5 Sonnet on Vertex AI ; June 20, 2024: newly released model now generally available on Vertex AI
  25. [25] Computer use, Claude 3.5 Sonnet, and Claude 3.5 Haiku — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  26. [26] Claude 3.7 Sonnet and Claude Code — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  27. [27] Introducing the next generation of Claude — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  28. [28] Claude Opus 4.1 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  29. [29] Introducing Claude Opus 4.5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  30. [30] OpenAI — introducing gpt 5 3 codex ; Publication date and availability section; Parameter counts not disclosed in launch article
  31. [31] Gemini 3: Introducing the latest Gemini AI model from Google ; Publication date and availability section; Official launch article; parameter counts not disclosed
  32. [32] Gemini 3.1 Pro: Announcing our latest Gemini AI model ; Publication date and availability section; Official launch article; parameter counts not disclosed
  33. [33] OpenAI — gpt 4 research ; Publication date and availability section; Parameter counts not disclosed in launch article
  34. [34] OpenAI DevDay ; Dated launch: GPT-4 Turbo preview available today
  35. [35] OpenAI — hello gpt 4o ; Publication date and availability section; Parameter counts not disclosed in launch article
  36. [36] OpenAI — gpt 5 1 codex max ; Publication date and availability section; Parameter counts not disclosed in launch article
  37. [37] OpenAI — introducing gpt 5 2 ; Publication date and availability section; Parameter counts not disclosed in launch article
  38. [38] OpenAI — introducing chatgpt pro ; Publication date and availability section; Parameter counts not disclosed in launch article
  39. [39] OpenAI — introducing openai o1 preview ; Publication date and availability section; Parameter counts not disclosed in launch article
  40. [40] OpenAI — introducing o3 and o4 mini ; Publication date and availability section; Parameter counts not disclosed in launch article
  41. [41] Introducing Claude Opus 4.8 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  42. [42] Introducing Claude Opus 5.5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  43. [43] Introducing Claude Sonnet 5.5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  44. [44] Introducing Claude Haiku 5.5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
  45. [45] deepseek-v4-flash developer model card ; Introduction; model comparison table
  46. [46] Meta — muse-spark ; Dated launch and explicit availability; Parameter counts not disclosed in reviewed developer source
  47. [47] Meta — muse-wonder ; Dated launch and explicit availability; Parameter counts not disclosed in reviewed developer source
  48. [48] Meta — muse-spark-1-1 ; Dated launch and explicit availability; Parameter counts not disclosed in reviewed developer source
  49. [49] Mistral Large 4 ; Dated Public Preview header; Model description: 1.05T total; 52B active; 1.6B vision encoder separately
  50. [50] qwen3-8-2-4t-a95b developer model card ; Model Overview: Number of Parameters
  51. [51] qwen3-8-flash-next developer model card ; Model Overview: Number of Parameters
  52. [52] glm-5 developer model card ; Introduction: model scaling
  53. [53] AgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges ; Official leaderboard: Level-0 Pass@3; arrays QR (web), $R (post), version map ez; routes-B-0Mubrq.js::QR/0/level0/pass3; routes-B-0Mubrq.js::QR/1/level0/pass3
  54. [54] AgentCyberRange: Evaluating autonomous cyber capabilities over long-horizon tasks ; Research introduction: Fudan–Whitzard team
  55. [55] Abhishek Jindal, Dmitry Kalashnikov 等. Can AI Perceive Physical Danger and Intervene?. ASIMOV v2 / arXiv:2509.21651 ; Project page author affiliation list
  56. [56] Andy K. Zhang, Neil Perry 等. Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models. 2025-04-12. v4 [arXiv]; Author affiliations on the original paper title page; Table 2
  57. [57] Zhun Wang, Tianneng Shi 等. CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale. 2026-03-24. v3 [arXiv]; Author affiliations on the original paper title page
  58. [58] Alexander Meinke, Bronson Schoen 等. Frontier Models are Capable of In-context Scheming. 2025-01-14. v2 [arXiv]; Author affiliations on the original paper title page
  59. [59] Linus Folkerts, Will Payne 等. Measuring AI Agents’ Progress on Multi-Step Cyber Attack Scenarios. Research page checked 2026-10-25 ; Official publication masthead and article byline
  60. [60] Thomas Kwa, Ben West 等. Task-Completion Time Horizons of Frontier AI Models. 2026-09-08
  61. [61] Thomas Kwa, Ben West 等. Measuring AI Ability to Complete Long Software Tasks ; Author affiliations on the original paper title page
  62. [62] Hjalmar Wijk, Tao Lin 等. RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts. 2025-05-27. v2 [arXiv]; Author affiliations on the original paper title page
  63. [63] Sid Black, Asa Cooper Stickland 等. RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents. 2025-05-05. v2 [arXiv]; Author affiliations on the original paper title page; Table 2
  64. [64] Xudong Pan (潘旭东), Jiarun Dai (戴嘉润) 等. Frontier AI systems have surpassed the self-replicating red line. 2024-12-09. v1 ; Author affiliations on the original paper title page; §1; §3.1
  65. [65] Rudolf Laine, Bilal Chughtai 等. Me, Myself, and AI: The Situational Awareness Dataset for LLMs. 2024-07-15 article; paper linked
  66. [66] Rudolf Laine, Bilal Chughtai 等. Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs ; Author affiliations on the original paper title page
  67. [67] Mary Phuong, Roland S. Zimmermann 等. Evaluating Frontier Models for Stealth and Situational Awareness. 2025-07-03. v4 [arXiv]; Author affiliations on the original paper title page; Table 1
  68. [68] Nathaniel Li ∗, Alexander Pan ∗ 等. The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning. 2024-05-15. v7 [arXiv]; Author affiliations on the original paper title page; Table 1
  69. [69] Yichen Wu, Qianqian Gao 等. OpenDeception: Learning Deception and Trust in Human-AI Interaction via Multi-Agent Simulation. 2026-02-06. v3 [arXiv]; Author affiliations on the original paper title page; Table 1
  70. [70] Olli Järviniemi, Oliver Makins 等. Propensity Inference: Environmental Contributors to LLM Behaviour. Research page checked 2026-10-25
  71. [71] Olli Järviniemi, Oliver Makins 等. Propensity Inference: Environmental Contributors to Unsanctioned LLM Behaviour ; Author affiliations on the original paper title page
  72. [72] Jeremy Schlatter, Benjamin Weinstein-Raun 等. Shutdown resistance in reasoning models. 2025-07-05 article; later TMLR version separate
  73. [73] Jeremy Schlatter, Benjamin Weinstein-Raun 等. Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs ; Author affiliations on the original paper title page
  74. [74] Edoardo Debenedetti, Jie Zhang 等. AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. 2024-11-24. v3 ; Author affiliations on the original paper title page; Appendix C, Table 3
  75. [75] Maksym Andriushchenko, Alexandra Souly 等. AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents. 2025-04-18. v3 [arXiv]; Author affiliations on the original paper title page; Table 2
  76. [76] Mantas Mazeika, Long Phan 等. HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal. 2024-02-27. v2 [arXiv]; Author affiliations on the original paper title page
  77. [77] Investigating unintended model actions in our evaluations and internal use. 2026-10-09 article ; Official publication masthead and article byline
  78. [78] Artem Petrov, Sergey Koldyba 等. Shutdown Resistance in Large Language Models, on robots!. 网页 2026-02-12 / 报告 2026-02-11 [PDF]; Authors; Technical Report (PDF), title page; §2–4; Figure 2 (p. 3)
  79. [79] Alexandra Souly ∗, Qingyuan Lu ∗ 等. A StrongREJECT for Empty Jailbreaks. 2024-08-27. v2 [arXiv]; Author affiliations on the original paper title page
  80. [80] Hwiwon Lee ∗ Jiawei Liu  Dongjun Kim  Wubing Xia  Ziqi Zhang  Chunqiu Steven Xia  Lingming Zhang ∗. SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?. 2026-07-20. v2 [arXiv]; Author affiliations on the original paper title page
  81. [81] Rahul Marchand, Art O Cathain 等. Quantifying Frontier LLM Capabilities for Container Sandbox Escape. AISI research page checked 2026-10-25
  82. [82] Rahul Marchand, Art O Cathain 等. Quantifying Frontier LLM Capabilities for Container Sandbox Escape ; Author affiliations on the original paper title page
  83. [83] Starace, Giulio, Jaffe, Oliver 等. PaperBench: Evaluating AI's Ability to Replicate AI Research. 2025-04-07. v3 [arXiv]; Author affiliations on the original paper title page; Table 4
  84. [84] Ryan Greenblatt, † Carson Denison † † footnotemark: 等. Alignment faking in large language models. 2024-12-18
  85. [85] Ryan Greenblatt, † Carson Denison † † footnotemark: 等. Alignment faking in large language models ; Author affiliations on the original paper title page
  86. [86] Agentic misalignment: How LLMs could be insider threats. 2025-06-20 ; Official publication masthead and article byline
  87. [87] SHADE-Arena: Evaluating sabotage and monitoring in LLM agents. 2025-06-16 ; Official publication masthead and article byline
  88. [88] Jasper Götting a, Pedro Medeiros a 等. Virology Capabilities Test (VCT): A Multimodal Virology Q&A Benchmark. 2025-04-29. v2 [arXiv]; Author affiliations on the original paper title page; Table 1
  89. [89] Introducing VCT-v2 — the updated Virology Capabilities Test. 2026-09-11 ; Official publication masthead and article byline
  90. [90] Eleanor M. Marshall, Pedro Medeiros 等. BioTIER: A Refusal Benchmark for Targeted Biological Risk Mitigation. 2026-07-16. v1 [arXiv]; Author affiliations on the original paper title page
  91. [91] Richard Ren, Arunim Agarwal 等. The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems. 2026-01-05. v3 [arXiv]; Author affiliations on the original paper title page; Table 3
  92. [92] Matthew Kowal, Jasper Timm 等. It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics. 2026-02-15. v4 [arXiv]; Author affiliations on the original paper title page
  93. [93] Large language models can effectively convince people to believe conspiracies. 2026-01-08 ; Official publication masthead and article byline
  94. [94] Hwiwon Lee  Ziqi Zhang  Hanxiao Lu † Lingming Zhang. SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks. 2025-06-13. v2 [arXiv]; Author affiliations on the original paper title page
  95. [95] Seunghyun Lee, David Brumley. ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents. 2026-05-13. v1 [arXiv]; Author affiliations on the original paper title page
  96. [96] Zhun Wang, Nico Schiller 等. ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?. 2026-05-13
  97. [97] Zhun Wang, Nico Schiller 等. ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? ; Author affiliations on the original paper title page
  98. [98] Yuxuan Zhu, Antony Kellermann 等. CVE-Bench: A Benchmark for AI Agents' Ability to Exploit Real-World Web Application Vulnerabilities. 2025-03-21. v4 [arXiv]; Author affiliations on the original paper title page
  99. [99] Jeremy Spence, Nicholas Assaderaghi 等. The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark. 2026-08-11. v2 [arXiv]; Author affiliations on the original paper title page
  100. [100] Assessing Claude Opus 5.5 Against Offensive Security Benchmarks. 2026-09-22 ; Official publication masthead and article byline
  101. [101] Assessing GPT-6 Astra: FrontierCyber Measures a Sharp Increase in Cyber Capability. 2026-09-03 ; Official publication masthead and article byline
  102. [102] Introducing SOLVE+: Scoring Cyber Scenarios (Version 0.5). 2026-09-03 ; Official publication masthead and article byline
  103. [103] Introduction ; Official publication masthead and article byline
  104. [104] Lauren Deason, Adam Bali 等. CyberSOCEval: Benchmarking LLMs Capabilities for Malware Analysis and Threat Intelligence Reasoning. v1
  105. [105] CyberSOCEval Benchmark Data ; Original project README: joint project between CrowdStrike and Meta
  106. [106] CAISI’s Assessment of Z.ai’s GLM-5.3 Cyber Capabilities. 2026-09-17 ; Official publication masthead and article byline
  107. [107] David Rein, Joel Becker 等. HCAST: Human-Calibrated Autonomy Software Tasks. 2025-03-21. v1 [arXiv]; Author affiliations on the original paper title page
  108. [108] MirrorCode: Evidence that AI can already do some weeks-long coding tasks. 2026-04-10 ; Official publication masthead and article byline
  109. [109] Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT. 2026-07-21 ; Official publication masthead and article byline
  110. [110] Chan Jun Shern, Neil Chowdhury 等. MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering. 2024-10-09. v6 [arXiv]; Author affiliations on the original paper title page; Table 2
  111. [111] Ben Rank ∗, Hardik Bhatnagar ∗ 等. PostTrainBench: Can LLM Agents Automate LLM Post-Training?. 2026-03-09. v2 [arXiv]; Author affiliations on the original paper title page; Table 1
  112. [112] Mary Phuong, Matthew Aitchison 等. Evaluating Frontier Models for Dangerous Capabilities. 2024-03-20. v2
  113. [113] Evaluating Frontier Models for Dangerous Capabilities. 2024-03-21 ; Official research publication and author list
  114. [114] Yihe Fan (范一禾), Minyuan Luo (罗铭源) 等. Large language model-powered AI systems achieve self-replication with no human intervention. 2025-03-14. v2 [arXiv]; Author affiliations on the original paper title page
  115. [115] Air, Alena, Reworr 等. Language Models Can Autonomously Hack and Self-Replicate. 2026-05-07. v1
  116. [116] Language Models Can Autonomously Hack and Self-Replicate. 2026-05-07 ; Official Palisade research publication matching the paper
  117. [117] Agentic Self-Modification in Open-Weights Systems. 2026-09-16 ; Official publication masthead and article byline
  118. [118] Alexander Pan, Jun Shern Chan 等. Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark. 2023-04-06. v4 [arXiv]; Author affiliations on the original paper title page
  119. [119] Jan Betley, Daniel Tan 等. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. 2025-02-25
  120. [120] Jan Betley, Daniel Tan 等. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs 1 This paper contains model-generated content that might be offensive. 1 ; Author affiliations on the original paper title page
  121. [121] Mia Taylor, James Chua 等. School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs. 2025-08-25
  122. [122] Mia Taylor, James Chua 等. School of reward hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs ; Author affiliations on the original paper title page
  123. [123] Jan Dubiński, Jan Betley 等. Conditional Misalignment: Common Interventions Can Hide Emergent Misalignment Behind Contextual Triggers. 2026-04-28
  124. [124] Jan Dubiński, Jan Betley 等. Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers ; Author affiliations on the original paper title page
  125. [125] From shortcuts to sabotage : Natural emergent misalignment from reward hacking. 2025-11-21 ; Official publication masthead and article byline
  126. [126] Axel Højmark, Jérémy Scheurer 等. Measuring Reward-Seeking via Contrastive Belief Updates. 2026-07-21. v1 [arXiv]; Author affiliations on the original paper title page
  127. [127] Robert Kirk, Alexandra Souly 等. Evaluating whether AI models would sabotage AI safety research. 2026-04-27 ; Official publication masthead and article byline
  128. [128] Teun van der Weij, MATS 等. AI Sandbagging: Language Models can Strategically Underperform on Evaluations. 2024-06-11. v4 [arXiv]; Author affiliations on the original paper title page
  129. [129] Evan Hubinger, Carson Denison † † footnotemark: , Jesse Mu † † footnotemark: , Mike Lambert † † footnotemark: , Meg Tong, Monte MacDiarmid,Tamera Lanham, Daniel M. Ziegler, Tim Maxwell, Newton ChengAdam Jermyn, Amanda Askell, Ansh Radhakrishnan, Cem Anil, David Duvenaud, Deep Ganguli 等. Sleeper Agents : Training deceptive LLMs that persist through safety training. 2024-01-14
  130. [130] Evan Hubinger, Carson Denison † † footnotemark: , Jesse Mu † † footnotemark: , Mike Lambert † † footnotemark: , Meg Tong, Monte MacDiarmid,Tamera Lanham, Daniel M. Ziegler, Tim Maxwell, Newton ChengAdam Jermyn, Amanda Askell, Ansh Radhakrishnan, Cem Anil, David Duvenaud, Deep Ganguli 等. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training ; Author affiliations on the original paper title page
  131. [131] Alex Cloud, Minh Le 等. Subliminal Learning: Language models transmit behavioral traits via hidden signals in data. 2025-07-20 ; Official publication masthead and article byline
  132. [132] Ryan Greenblatt ∗ Buck Shlegeris      Kshitij Sachan      Fabien Roger Redwood Research. AI Control: Improving Safety Despite Intentional Subversion. 2023-12-12. v5 [arXiv]; Author affiliations on the original paper title page
  133. [133] Aryan Bhatt, Cody Rushing 等. Ctrl-Z: Controlling AI Agents via Resampling. 2025-04-14. v1 [arXiv]; Author affiliations on the original paper title page
  134. [134] Adam Kaufman, James Lucassen. BashArena: A Control Setting for Highly Privileged AI Agents. 2025-12-17. v1 [arXiv]; Author affiliations on the original paper title page
  135. [135] Oliver Makins, Orazio Angelini 等. Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors. 2026-07-08
  136. [136] Oliver Makins, Orazio Angelini 等. Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors ; Author affiliations on the original paper title page
  137. [137] Stress Testing Deliberative Alignment for Anti-Scheming Training. 2025-09-17 ; Official publication masthead and article byline
  138. [138] MALT: A Dataset of Natural and Prompted Behaviors That Threaten Eval Integrity. 2025-10-14 ; Official publication masthead and article byline
  139. [139] Early work on monitorability evaluations. 2026-01-22 ; Official publication masthead and article byline
  140. [140] Jordan Taylor, Sid Black Dillon Bowen Thomas Read Satvik Golechha 等. Auditing Games for Sandbagging. 2025-12-08. v1 [arXiv]; Author affiliations on the original paper title page
  141. [141] Alan Cooney, David Africa 等. "Did you lie?": Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms. 2026-06-17 ; Official publication masthead and article byline
  142. [142] The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes. 2026-02-16 ; Official publication masthead and article byline
  143. [143] Saad Hossain, Tom Tseng 等. TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering. 2026-02-06. v2 [arXiv]; Author affiliations on the original paper title page
  144. [144] Automated researchers can reliably mitigate alignment failures. 2026-08-28 ; Official publication masthead and article byline
  145. [145] We need 3rd party Training-Run Evaluations. 2026-07-05 ; Official publication masthead and article byline
  146. [146] Reasoning models don't always say what they think. 2025-04-03 ; Official publication masthead and article byline
  147. [147] Bowen Baker Joost Huizinga *† Leo Gao † Zehao Dou † Melody Y. Guan † Aleksander Madry † Wojciech Zaremba † Jakub Pachocki † David Farhi. Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation. 2025-03-14. v1 [arXiv]; Author affiliations on the original paper title page
  148. [148] Qiusi Zhan, Zhixiang Liang 等. InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents. 2024-03-05. v3 [arXiv]; Author affiliations on the original paper title page
  149. [149] Zhexin Zhang 1 1 1 Equal contribution., Shiyao Cui 1 1 1 Equal contribution. 等. Agent-SafetyBench: Evaluating the Safety of LLM Agents. 2024-12-19. v2 [arXiv]; Author affiliations on the original paper title page
  150. [150] Petri : An open-source auditing tool to accelerate AI safety research. 2025-10-06 ; Official publication masthead and article byline
  151. [151] Introducing Bloom : an open source tool for automated behavioral evaluations. 2025-12-19 ; Official publication masthead and article byline
  152. [152] Stephanie Lin, Jacob Hilton 等. TruthfulQA: Measuring how models mimic human falsehoods. 2021-09-08
  153. [153] Stephanie Lin, Jacob Hilton 等. TruthfulQA: Measuring How Models Mimic Human Falsehoods ; Author affiliations on the original paper title page
  154. [154] Jan Betley, Johannes Treutlein † † footnotemark: 等. Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values. 2026-07-15
  155. [155] Jan Betley, Johannes Treutlein † † footnotemark: 等. Value Leakage: An LLM’s Answers Are Silently Shaped by Its Own Values ; Author affiliations on the original paper title page
  156. [156] Insights into AI Agent Security from a Large-Scale Red-Teaming Competition. 2026-03-23 ; Official publication masthead and article byline
  157. [157] Jon M Laurent, Albert Bou 等. LABBench2: An Improved Benchmark for AI Systems Performing Biology Research. 2026-02-04. v2 [arXiv]; Author affiliations on the original paper title page
  158. [158] Christina Q. Knight ∗, Kaustubh Deshpande ⋄ 等. FORTRESS: Frontier Risk Evaluation for National Security and Public Safety. 2025-06-17. v2 [arXiv]; Author affiliations on the original paper title page
  159. [159] Dev, Sunishchal, Teague, Charles 等. Toward Comprehensive Benchmarking of the Biological Knowledge of Frontier Large Language Models. 2025-11-25 ; Official publication masthead and article byline
  160. [160] Brent, Roger, McKelvey, Greg, Jr.. Contemporary Foundation AI Models Increase Biological Weapons Risk. 2025-12-31 ; Official publication masthead and article byline
  161. [161] Measuring tactical intelligence targeting and conventional weapons capabilities of AI models. 2026-09-10 ; Official publication masthead and article byline
  162. [162] Laurence Wroe. Auditing FORTRESS: A Multi-Judge Panel Doesn’t Guarantee Robustness ; Article byline, 18 September 2026; What is FORTRESS?; audit methodology
  163. [163] Cheating On AI Agent Evaluations. 2025-12-02 ; Official publication masthead and article byline
  164. [164] Joshua Fonseca Rivera, Neil Shah 等. Item Response Theory for AI Safety. 2026-08-05 ; Official publication masthead and article byline
  165. [165] Toby D. Pilditch. Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations. 2026-08-14
  166. [166] Toby D. Pilditch. Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations ; Author affiliations on the original paper title page
  167. [167] Frontier Risk Report (February to March 2026). 2026-05-19 ; Executive summary and guide to the report; Pilot process
  168. [168] Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident. 2026-08-26 ; Official publication masthead and article byline
  169. [169] An alignment assessment of recent cybersecurity incidents. 2026-09-09 ; Official publication masthead and article byline
  170. [170] Detecting and countering misuse of AI: September 2026 ; Official publication masthead and article byline
  171. [171] GLM-5.2 Risk Evaluation Report. 2026-08-02
  172. [172] GLM-5.2 Risk Evaluation Report. 2026-08-02. v1.1 · revised 26 August 2026 ; Report masthead and stated evaluation methodology; §2.2; Figure 3; Appendix A.7; Table A.3, pp. 41–42
  173. [173] Andy K. Zhang, Joey Ji 等. BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems. 2025-05-21. v3 [arXiv]; Author affiliations on the original paper title page; Table 1
  174. [174] Kyle O’Brien, Stephen Casper 等. Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs. 2025-08-08. v2 [arXiv]; Author affiliations on the original paper title page
  175. [175] Miles Wang, Micah Carroll 等. Monitoring Monitorability. 2025-12-20. v1 [arXiv]; Paper author block and correspondence at openai.com
  176. [176] Tomek Korbak * UK AI Security Institute Mikita Balesni * Apollo Research Elizabeth Barnes METR Yoshua Bengio University of Montreal & Mila Joe Benton Anthropic Joseph Bloom UK AI Security Institute Mark Chen OpenAI Alan Cooney UK AI Security Institute Allan Dafoe Google DeepMind Anca Dragan Google DeepMind Scott Emmons Google DeepMind Owain Evans Truthful AI & UC Berkeley David Farhi OpenAI Ryan Greenblatt Redwood Research Dan Hendrycks Center for AI Safety Marius Hobbhahn Apollo Research Evan Hubinger Anthropic Geoffrey Irving UK AI Security Institute Erik Jenner Google DeepMind Daniel Kokotajlo AI Futures Project Victoria Krakovna Google DeepMind Shane Legg Google DeepMind David Lindner Google DeepMind David Luan Amazon Aleksander Mądry OpenAI Julian Michael Scale AI Neel Nanda Google DeepMind Dave Orr Google DeepMind Jakub Pachocki OpenAI Ethan Perez Anthropic Mary Phuong Google DeepMind Fabien Roger Anthropic Joshua Saxe Meta Buck Shlegeris Redwood Research Martín Soto UK AI Security Institute Eric Steinberger Magic Jasmine Wang UK AI Security Institute Wojciech Zaremba OpenAI Bowen Baker † OpenAI Rohin Shah † Google DeepMind Vlad Mikulik † Anthropic. Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. 2025-07-15. v2 [arXiv]; Author affiliations on the original paper title page
  177. [177] Han Wang, Yifan Sun 等. MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models. 2026-03-30. v2 [arXiv]; Author affiliations on the original paper title page
  178. [178] Your Name. Catastrophic Cyber Capabilities Benchmark (3CB): Robustly Evaluating LLM Agent Cyber Offense Capabilities. 2024-10-10. v2
  179. [179] ‘3cb’: The Catastrophic Cyber Capabilities Benchmark. 2024-10-30 ; Original institutional announcement: Apart Research paper
  180. [180] Yufei He, Yuexin Li 等. Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?. 2025-02-16. v1 [arXiv]; Author affiliations on the original paper title page; Table 1
  181. [181] Zhiqiang Wang, Yichao Gao 等. MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP Servers. 2025-08-19. v2
  182. [182] Zhiqiang Wang, Yichao Gao 等. MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP Servers — AAAI-26 paper. AAAI-26 proceedings, pp. 35811–35819 ; AAAI-26 original paper, p. 1, author affiliations 1–2
  183. [183] Sijun Tan, Siyuan Zhuang 等. JudgeBench: A Benchmark for Evaluating LLM-based Judges. 2024-10-16. v2 [arXiv]; Author affiliations on the original paper title page; Table 2
  184. [184] Aaron Rose, Carissa Cullen 等. Detecting Multi-Agent Collusion Through Multi-Agent Interpretability. 2026-04-01. v2 [arXiv]; Author affiliations on the original paper title page
  185. [185] Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations. 2026-10-07 ; Official publication masthead and article byline
  186. [186] Toby Shevlane, Sebastian Farquhar 等. Model evaluation for extreme risks. 2023-05-24. v2 [arXiv]; Author affiliations on the original paper title page
  187. [187] Melody Guan, Yuxiang Xie 等. Beyond g-mean²: complementary metrics for monitorability. 2026-10-20 ; Official publication masthead and article byline
  188. [188] Prefill Awareness in Large Language Models. 2026-06-10 ; Official publication masthead and article byline
  189. [189] Consistency Training Can Entrench Misalignment. 2026-06-02 ; Official publication masthead and article byline
  190. [190] Loss of Oversight: How AI systems may become harder to audit, monitor, and investigate. 2026-05-21 ; Official publication masthead and article byline
  191. [191] A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring. 2026-04-29 ; Official publication masthead and article byline
  192. [192] Seven simple steps for log analysis in AI systems. 2026-02-25 ; Official publication masthead and article byline
  193. [193] Practical challenges of control monitoring in frontier AI deployments. 2025-12-15 ; Official publication masthead and article byline
  194. [194] Async control: Stress-testing asynchronous control measures for LLM agents. 2025-12-15 ; Official publication masthead and article byline
  195. [195] Does self-evaluation enable wireheading in language models?. 2025-12-01 ; Official publication masthead and article byline
  196. [196] Open technical problems in open-weight AI model risk management. 2025-10-26 ; Official publication masthead and article byline
  197. [197] Alexandra Souly, Javier Rando 等. Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples. 2025-10-08
  198. [198] Alexandra Souly, Javier Rando 等. Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples ; Author affiliations on the original paper title page
  199. [199] Daniel Tan, Anders Woodruff 等. Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time. 2025-10-05
  200. [200] Daniel Tan, Anders Woodruff 等. Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time ; Author affiliations on the original paper title page
  201. [201] An example safety case for safeguards against misuse. 2025-06-05 ; Official publication masthead and article byline
  202. [202] Existing Large Language Model unlearning evaluations are inconclusive. 2025-05-31 ; Official publication masthead and article byline
  203. [203] Safety Cases: A scalable approach to Frontier AI safety. 2025-02-10 ; Official publication masthead and article byline
  204. [204] Principles for evaluating misuse safeguards of frontier AI systems. 2025-02-04 ; Official publication masthead and article byline
  205. [205] A sketch of an AI control safety case. 2025-01-28
  206. [206] A sketch of an AI control safety case ; Author affiliations on the original paper title page
  207. [207] Embedded Evaluators are necessary for meaningful external testing. 2026-09-29 ; Official publication masthead and article byline
  208. [208] Demonstrating specification gaming in reasoning models. 2025-02-19 ; Official publication masthead and article byline
  209. [209] LLM Honeypot: an early warning system for autonomous hacking. 2024-10-17 ; Official publication masthead and article byline
  210. [210] Jan Dubiński, Anna Sztyber-Betley 等. Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors. 2026-10-07 ; Official publication masthead and article byline
  211. [211] Adam Karvonen, James Chua 等. Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers. 2025-12-19 ; Official publication masthead and article byline
  212. [212] The Rogue Replication Threat Model. 2024-11-12 ; Official publication masthead and article byline
  213. [213] Guidelines for capability elicitation. 2024-03-15 ; Official publication masthead and article byline
  214. [214] AbstentionBench. 2025-06-10
  215. [215] DeceptionBench. 2025-10-17
  216. [216] PersuSafety. 2025-04-14
  217. [217] ImpossibleBench. 2025-10-23
  218. [218] EvalAwareBench
  219. [219] TheAgentCompany. 2024-12-18
  220. [220] LAB-Bench. 2024-07-14
  221. [221] CharXiv
  222. [222] BioMysteryBench
  223. [223] EVMbench
  224. [224] SWE-Bench Pro
  225. [225] CTF-Archive-Diamond. 2026-05-01. 1 May 2026 ; Report masthead and stated evaluation methodology; Capability Results; Figure 2; *** footnote; Capability Results; Figure 2
  226. [226] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; Report masthead and stated evaluation methodology; §1.1.1; Figure 1; §1.1.2; Figure 2
  227. [227] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; Report masthead and stated evaluation methodology; PDF p. 8; Cybench unguided; pass@1; PDF p. 9; EVMbench; Detect
  228. [228] Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks. 2026-07-03. 3 July 2026 ; Report masthead and stated evaluation methodology; §1; Figure 1; §1; Figure 2
  229. [229] Evaluation of DeepSeek AI Models. 2025-09-30. 30 September 2025 ; Report masthead and stated evaluation methodology; p. 24, Table 4.1, row CVE-Bench, column OpenAI GPT-5; p. 24, Table 4.1, row CVE-Bench, column Anthropic Opus 4
  230. [230] Red-Teaming OpenAI’s Internal Agent Monitoring Systems. 2026-10-22 ; Public findings and footnotes; text lines 00058
  231. [231] Incident Report: unsanctioned agent behaviour during cyber testing ; What we found; evaluation setup; text lines 00037,00038
  232. [232] Scaling Misuse Detection with Self-Assessment. 2026-10-16 ; Results; limitations; text lines 00008,00014,00060
  233. [233] TraceProbe: An Execution-Based Benchmark for Diagnosing Safety Failures in LLM Agents
  234. [234] Cheng Yu, Severin Engelmann 等. How Should AI Safety Benchmarks Benchmark Safety?
  235. [235] Kobi Hackenburg, Caroline Wagner 等. AI systems out-persuade expert humans. 2026-06-15
  236. [236] Ibrahim, Lujain, Hafner, Franziska Sofia 等. Training language models to be warm can reduce accuracy and increase sycophancy. 2026-04-29
  237. [237] Edward Sun, Christopher Leet 等. Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control. 2026-10-03
  238. [238] Arnav Balaji, Arpit Bahety 等. OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation. 2026-06-30
  239. [239] Mingyang Lyu, Yinqian Sun 等. ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models
  240. [241] Long-Form Tasks: A Methodology for Evaluating Scientific Assistants ; Executive summary; evaluation trade-offs; text lines 00020,00035
  241. [242] Impact of modeling assumptions on time horizon results. 2026-03-20 ; Summary; model correction; uncertainty table; text lines 00043,00069,00240
  242. [243] Clarifying limitations of time horizon. 2026-01-22 ; Research note and limitations; text lines 00072
  243. [244] HiBayES: Improving LLM evaluation with hierarchical Bayesian modelling ; Official methods article; text lines 00001,00012,00013,00018,00021,00024,00025,00027
  244. [245] Shiv Muthupandiyan, Tolga H. Dur 等. ReproBAIT: Can coding agents build their own biological AI tools?. 2026-10-14
  245. [246] Aaron Maiwald, Roderick McNeill 等. The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models. 2026-10-19
  246. [247] Open Problems in Frontier AI Risk Management
  247. [248] Sabotage evaluations for frontier models. 2024-10-18
  248. [249] How we monitor internal coding agents for misalignment. 2026-03-19
  249. [250] Findings from a pilot Anthropic–OpenAI alignment evaluation exercise: OpenAI Safety Tests. 2025-08-27
  250. [251] Anthropic's Pilot Sabotage Risk Report. 2025-10-28
  251. [252] TASTE: Can AI Models Judge AI Safety Research Proposals?. 2026-08-28
  252. [253] Modular Pretraining Enables Access Control. 2026-07-08
  253. [254] Gram: Assessing sabotage propensities via automated alignment auditing. 2026-05-28
  254. [255] Strengthening our Frontier Safety Framework. 2025-09-22
  255. [256] How confessions can keep language models honest. 2025-12-03
  256. [257] Reasoning models struggle to control their chains of thought, and that’s good. 2026-03-05
  257. [258] Strengthening our safety ecosystem with external testing. 2025-11-19
  258. [259] Research Note: Our scheming precursor evals had limited predictive power for our in-context scheming evals. 2025-07-03
  259. [260] Introducing LinuxArena. 2026-04-20 ; Article image 1; label Claude Opus 4.6; Main task performance; figure reports n=400 each; task count versus repeated runs not separately disclosed
  260. [261] Announcing the ARC White-Box Estimation Challenge — Alignment Research Center. 2026-06-02
  261. [262] Summary of METR's predeployment evaluation of Claude Opus 5.5. 2026-09-22
  262. [263] Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report (August 2026). 2026-10-20
  263. [264] AI systems could cover up misbehavior. 2026-10-06
  264. [265] Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming. 2026-10-16
  265. [266] Retrying vs Resampling in AI Control
  266. [267] Simulated Control Evaluations. 2026-10-24
  267. [268] Latent reasoning architectures would likely undermine CoT, our strongest oversight tool. 2026-09-23
  268. [269] Measuring Reward-Seeking via Contrastive Belief Updates. 2026-07-21
  269. [270] Towards embedded evaluations for scheming propensities. 2026-10-01
  270. [271] Logit ROCs: Monitor TPR is linear in FPR in logit space
  271. [272] Evaluating AI cyber capabilities with crowdsourced elicitation. 2025-05-26
  272. [273] BashArena and Control Setting Design. 2025-12-18 ; Article image1; main task performance; bar GPT-5.2; Article image1; main task performance; bar Claude Opus 4.5; Article image1; main task performance; bar Claude Sonnet 4.5
  273. [274] Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents. 2024-10-03 ; #S5.T5; row Gemma2-9B; condition DPI; column Attack success rate; #S5.T5; row Gemma2-9B; condition DPI; column Refusal rate; #S5.T5; row Gemma2-9B; condition IPI; column Attack success rate
  274. [275] SafeAgentBench: A Benchmark for Safe Task Planning of Embodied LLM Agents. 2024-12-17 ; #S4.T2; row GPT-4 / Lota-Bench; condition Detailed hazardous tasks; column Rejection ratio; #S4.T2; row GPT-4 / Lota-Bench; condition Detailed hazardous tasks; column Risk ratio (goal evaluator)
  275. [276] Identifying the Risks of LM Agents with an LM-Emulated Sandbox. 2023-09-25 ; #S5.T5; row GPT-4 / Basic; condition ToolEmu adversarial emulator / Basic; column Average safety score; #S5.T5; row GPT-4 / Basic; condition ToolEmu adversarial emulator / Basic; column Failure incidence
  276. [277] AI Agents That Matter. 2024-07-01
  277. [278] Measuring what Matters: Construct Validity in Large Language Model Benchmarks. 2025-11-03
  278. [279] AI Evaluation Should Require Standardized Item-Level Data Releases. 2026-02-27
  279. [280] Tensor Trust: Interpretable Prompt Injection Attacks from an Online Game. 2023-11-02 ; #A3.T1; row gpt-3.5-turbo-0613; column Hijacking robustness rate; condition Hijacking; #A3.T1; row gpt-3.5-turbo-0613; column Defense validity; condition Hijacking
  280. [281] AgentVigil: Generic Black-Box Red-teaming for Indirect Prompt Injection against LLM Agents. 2025-05-09
  281. [282] RedCode: Risky Code Execution and Generation Benchmark for Code Agents. 2024-11-12 ; #S4.T1; row GPT-4o / Base LLM; condition RedCode-Gen / Base LLM; column Accuracy; #S4.T1; row GPT-4o / Base LLM; condition RedCode-Gen / Base LLM; column Refusal rate
  282. [283] AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases. 2024-07-17
  283. [284] Measurement to Meaning: A Validity-Centered Framework for AI Evaluation. 2025-05-13
  284. [285] What Victims Take Away: Attack Reuse and Defense Revision after Prompt-Injection Attacks. 2026-10-27
  285. [286] evalci: A Python Library for Statistically Rigorous Comparison of Language Model Evaluations. 2026-07-05
  286. [287] Resolution Diagnostics for Paired LLM Evaluation. 2026-05-28
  287. [288] ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three Stages. 2024-02-16
  288. [289] R-Judge: Benchmarking Safety Risk Awareness for LLM Agents. 2024-01-18
  289. [290] METR Time Horizon 1.1 — source data ; results.claude_3_5_sonnet_20240620_inspect; results.claude_3_5_sonnet_20241022_inspect; results.claude_3_7_sonnet_inspect
  290. [291] Anthropic February 2026 Risk Report. 2026-07-08
  291. [292] Claude Opus 4.5 System Card. 2025-11-05. 2025-12-05
  292. [293] Claude Opus 4.6 System Card. 2026-02-06. 2026-03-06
  293. [294] Claude Opus 4.6 Sabotage Risk Report. 2026-02-10
  294. [295] Claude Sonnet 4.6 System Card. 2026-02-17. 2026-03-06
  295. [296] Alignment Risk Update: Claude Mythos Preview. 2026-04-07. 2026-04-10
  296. [297] Claude Opus 4.7 System Card. 2026-04-16
  297. [298] Claude Opus 4.8 System Card. 2026-05-28. 2026-06-17
  298. [299] Claude Sonnet 5 System Card. 2026-06-30. 2026-07-10
  299. [300] Anthropic August 2026 Risk Report. 2026-08-14
  300. [301] Claude Fable 5.1 and Mythos 5.1 System Card. 2026-09-01
  301. [302] Claude Opus 5.5 System Card. 2026-09-22
  302. [303] Claude Sonnet 5.5 System Card. 2026-09-28
  303. [304] Claude Haiku 5.5 System Card. 2026-10-07
  304. [305] DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  305. [306] Gemini 4 Argon Safety & Security Report
  306. [307] Gemini 4 Argon Model Card. 2026-10-13
  307. [308] Muse Glimmer Model Card
  308. [309] Muse Spark 1.2 Multimodal Evaluation Methodology
  309. [310] Muse Spark 1.3 Evaluation Methodology
  310. [311] Muse Spark Contemplating Safety & Preparedness Report
  311. [312] Muse Spark Open Evaluation Methodology
  312. [313] Muse Spark Open Model Card
  313. [314] Muse Spark Safety & Preparedness Report (Meta-hosted May 26 revision). 2026-05-26
  314. [315] Muse Wonder Model Card. 2026-10-23
  315. [316] Muse Spark Safety & Preparedness Report. 2026-05-14. 2026-08-24
  316. [317] Muse Spark 1.1 Evaluation Report. 2026-07-09
  317. [318] Shieldstral
  318. [319] Kimi K3 Technical Report
  319. [320] Update to GPT-5 System Card: GPT-5.2. 2025-12-11
  320. [321] GPT-5.4 Thinking System Card. 2026-03-05
  321. [322] GPT-5.5 System Card. 2026-04-23
  322. [323] GPT-5.6 System Card. 2026-07-09
  323. [324] GPT-6 Astra System Card. 2026-09-03. 2026-09-09
  324. [325] Addendum to GPT-6 Astra System Card: GPT-6.1 Sol. 2026-09-29
  325. [326] GPT-6 Sol and GPT-6 Luna: October 2026 update. 2026-10-07
  326. [327] Qwen3Guard Technical Report
  327. [328] On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability. 2026-08-26
  328. [329] Grok 4.7 Model Card. 2026-09-21
  329. [330] GLM-5.3-BF16 Model Card
  330. [331] GLM-5: from Vibe Coding to Agentic Engineering
  331. [332] CyberGym official leaderboard and machine-data extraction ; /level1/0/score_10; /level1/1/score_10; /level1/2/score_10
  332. [333] ExploitGym official leaderboard and benchmark page ; /results/0/on_target; /results/0/subRows/0/on_target; /results/1/on_target
  333. [334] SEC-bench Pro official leaderboard: selected 260617 and historical 260505 ; /0/results/0/score_modes/headline/score; /0/results/0/score_modes/completed/score; /0/results/1/score_modes/headline/score
  334. [335] Jack Payne, Jeremy Miller 等. Offensive Cybersecurity Time Horizons. 2026-04-02
  335. [336] Jack Payne, Sean Peters. GPT-5.5 Saturates Our Offensive Cybersecurity Time Horizons. 2026-05-27
  336. [337] Lyptus Cyber Task Horizons: aggregate model runs