风险评测 · 技术报告

Frontier Risk Report (February to March 2026)

NUWA 研究解读 · 独立复核待完成

首次发表: 2026-05-19来源核验: 2026-10-27

风险监测

原始来源 ↗引用本记录 ↓自主性 →欺骗 →

NUWA Research Note

研究问题研究设计与假设来源发现 NUWA 解读局限与相反证据研究与决策意义机构与角色原始资源

模型辅助整理 · 选读原始全文或官方文章中的相关章节；尚未完成逐章节审计。

研究速读

NUWA 解读：内部使用风险需同时审查组织访问、控制及模型能力。

研究问题

周期性机构级评估如何覆盖开发者内部AI使用风险？

研究设计与假设

2026年2月16日至3月16日试点涉及四家开发者，取得内部模型、原始推理及非公开组织信息，并经参与者审核披露。

来源发现

报告分开手段、动机和机会。

强编程能力并存判断可靠性限制，监测有效但有例外，模型分数不是完整风险证据。

NUWA 解读

NUWA 解读：内部使用风险需同时审查组织访问、控制及模型能力。

局限与相反证据

此次仅选读执行摘要、试点过程和部分监测结果，未完整审计长报告或保密证据。

研究与决策意义

模型建议：引用评估窗口，并区分公开观察、开发者信息及METR推断。

精确来源位置

Frontier Risk Report (February to March 2026) · Executive summary and guide to the reportSelected original resource https://metr.org/blog/2026-05-19-frontier-risk-report/; sha256:b0ebbde94fc97264cfac3378eb3794a8caba87ff4a51d35f854a71a565696d8a; archived source capture; exact capture time not independently established
Frontier Risk Report (February to March 2026) · Pilot processSelected original resource https://metr.org/blog/2026-05-19-frontier-risk-report/; sha256:b0ebbde94fc97264cfac3378eb3794a8caba87ff4a51d35f854a71a565696d8a; archived source capture; exact capture time not independently established
Frontier Risk Report (February to March 2026) · Opportunity: Monitors caught many harmful actions but there were exceptions and workaroundsSelected original resource https://metr.org/blog/2026-05-19-frontier-risk-report/; sha256:b0ebbde94fc97264cfac3378eb3794a8caba87ff4a51d35f854a71a565696d8a; archived source capture; exact capture time not independently established
Frontier Risk Report (February to March 2026) · No-reasoning and visible reasoning resultsSelected original resource https://metr.org/blog/2026-05-19-frontier-risk-report/; sha256:b0ebbde94fc97264cfac3378eb3794a8caba87ff4a51d35f854a71a565696d8a; archived source capture; exact capture time not independently established

来源核查、数据提取、编辑批准与实验复现分别记录。方法说明 →

机构与角色

评测机构 METR

发布机构 METR

引用与复用

引用研究发现时请引用原始出版物；引用 NUWA 的整理与解释时，单独引用本站记录及对应快照。

BibTeX ↓RIS ↓记录与来源元数据 · JSON ↓阅读覆盖 · JSON 提交更正

原始出版物 · BibTeX

@misc{metr-frontier-risk-report,
  title = {Frontier Risk Report (February to March 2026)},
  url = {https://metr.org/blog/2026-05-19-frontier-risk-report/},
  year = {2026},
  date = {2026-05-19},
  version = {Source capture 2026-10-25},
  note = {Source capture 2026-10-25}
}

NUWA 记录 · BibTeX

@misc{metr-frontier-risk-report-nuwa,
  title = {Frontier Risk Report (February to March 2026) — NUWA research record},
  url = {https://whitzard.tech/nuwa/whitzard-index/work/metr-frontier-risk-report/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--metr-frontier-risk-report.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--metr-frontier-risk-report.json.}
}

证据快照 2026-10-29.1 · 索引更新 · RSS

参考文献

[1] Frontier Risk Report (February to March 2026). 2026-05-19 ; Executive summary and guide to the report; Pilot process; Executive summary and guide to the report; Pilot process
[2] MALT: A Dataset of Natural and Prompted Behaviors That Threaten Eval Integrity. 2025-10-14
[3] the APPS dataset (Hendrycks et al. 2021)
[4] agents can construct encoded messages