风险发现 · 研究

Language Models Can Autonomously Hack and Self-Replicate

NUWA 研究解读 · 独立复核待完成

首次发表: 2026-05-07来源核验: 2026-10-25

原始来源 ↗引用本记录 ↓自主性 →网络安全 →

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联系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

该任务连接获取访问与部署可运行模型栈；API 模型复制的是另一模型的权重。

研究问题

接受指令的智能体能否同时完成访问获取与建立运行副本，副本又能否继续下一跳？

研究设计与假设

研究使用刻意设置脆弱点的目标，分别记录访问与复制里程碑；开放权重和 API 系统脚手架不同，链式成功要求更严格。

来源发现

副本能够响应推理请求，不等于能够启动下一跳。

API 模型实验部署 Qwen 权重，不能据此声称复制了专有模型权重。

所报运行副本率排除拒绝运行，需同时保留拒绝数与总运行数。

NUWA 解读

NUWA 解读：该流程弥补预先提供访问的评测范围缺口，但仍是受控能力展示。

局限与相反证据

目标难度较低且无纵深防御，脚手架和提示存在差异，限制真实部署外推与模型家族比较。

研究联系

NUWA 关联：扩展复制研究预先提供目标访问，RepliBench 分测前置能力；两者分数都不能与该流程互换。

研究与决策意义

NUWA 建议：记录载荷身份和逐跳判据，用于设计隔离评测，避免传播攻击操作。

精确来源位置

Language Models Can Autonomously Hack and Self-Replicate · 3 Experimental SetupSelected original resource https://arxiv.org/pdf/2605.06760; sha256:232c5659f34d6da1a3a556b4f0224a2d95214a40bf59d759cb85feeef1430b18; archived source capture; exact capture time not independently established
Language Models Can Autonomously Hack and Self-Replicate · 4.1 MilestonesSelected original resource https://arxiv.org/pdf/2605.06760; sha256:232c5659f34d6da1a3a556b4f0224a2d95214a40bf59d759cb85feeef1430b18; archived source capture; exact capture time not independently established
Language Models Can Autonomously Hack and Self-Replicate · 5 ResultsSelected original resource https://arxiv.org/pdf/2605.06760; sha256:232c5659f34d6da1a3a556b4f0224a2d95214a40bf59d759cb85feeef1430b18; archived source capture; exact capture time not independently established
Language Models Can Autonomously Hack and Self-Replicate · 7 LimitationsSelected original resource https://arxiv.org/pdf/2605.06760; sha256:232c5659f34d6da1a3a556b4f0224a2d95214a40bf59d759cb85feeef1430b18; 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

发布机构 Palisade Research

作者 (6)

Alena Air · Reworr · Nikolaj Kotov · Dmitrii Volkov · John Steidley · Jeffrey Ladish

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{palisade-network-replication,
  title = {Language Models Can Autonomously Hack and Self-Replicate},
  url = {https://arxiv.org/abs/2605.06760},
  author = {Alena Air and Reworr and Nikolaj Kotov and Dmitrii Volkov and John Steidley and Jeffrey Ladish},
  year = {2026},
  date = {2026-05-07},
  version = {arXiv:2605.06760v1},
  note = {arXiv:2605.06760v1}
}

NUWA 记录 · BibTeX

@misc{palisade-network-replication-nuwa,
  title = {Language Models Can Autonomously Hack and Self-Replicate — NUWA research record},
  url = {https://whitzard.tech/nuwa/whitzard-index/work/palisade-network-replication/},
  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--palisade-network-replication.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--palisade-network-replication.json.}
}

证据快照 2026-10-29.1 · 索引更新 · RSS

参考文献

[1] Air, Alena, Reworr 等. Language Models Can Autonomously Hack and Self-Replicate. 2026-05-07. v1 ; 3 Experimental Setup; 4.1 Milestones; 5 Results
[2] Language Models Can Autonomously Hack and Self-Replicate. 2026-05-07 ; Official Palisade research publication matching the paper