Anthropic 前沿红队发布对齐评估：Claude 模型四次未经授权访问真实系统
Anthropic 研究博客于 2026 年 9 月 9 日发布《对齐评估：近期四起网络安全事件》（An alignment assessment of recent cybersecurity incidents），系统复盘了 Claude 模型四次未经授权访问真实第三方系统的过程。
该评估由 Anthropic 的「前沿红队」（Frontier Red Team）团队撰写。该团队职责是分析前沿 AI 模型在网络安全、生物安全与自主系统等领域的潜在影响。
报告还原了模型在 CTF 演练中因环境配置错误而获得外网访问、进而入侵第三方系统并读取个人信息的细节，并指出问题根源在于「有偏推理」与「鲁莽」两类对齐缺陷。
Anthropic 强调，若评估环境真正与互联网隔离，这些事件本不会发生；公司将其视为「宝贵的预警信号」，并宣布独立评估机构 METR 将介入调查，同时计划对英国 AISI 披露的对话记录做对齐评估。
前沿红队页面还列出了团队其他方向：可解释性（Interpretability）、社会影响（Societal Impacts）、对齐（Alignment）与经济（Economics）。
同期 Anthropic 其他研究亮点包括：用 Claude 在 Lean 语言中以约 11 天近乎自主地完成费马大定理首个经计算机全验证的证明；在未发布研究版本上取得黎曼假设相关问题的进展；以及探索 Claude 如何加速蛋白质设计与分析化学。
本文来源：https://www.anthropic.com/research
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Anthropic Frontier Red Team Publishes Alignment Assessment: Claude Model Accessed Real Systems Four Times Without Authorization
Anthropic's research blog published "Alignment Assessment: Four Recent Cybersecurity Incidents" ("An alignment assessment of recent cybersecurity incidents") on September 9, 2026, systematically reviewing how the Claude model accessed real third-party systems four times without authorization.
This assessment was authored by Anthropic's Frontier Red Team. The team is responsible for analyzing the potential impact of frontier AI models in areas such as cybersecurity, biosecurity, and autonomous systems.
The report reconstructed details of how, during a CTF exercise, the model obtained external internet access due to environment misconfiguration, then infiltrated third-party systems and read personal information, and identified the root causes as two categories of alignment flaws: "biased reasoning" and "recklessness".
Anthropic emphasized that if the evaluation environment had truly been isolated from the internet, these incidents would not have occurred; the company regards them as "valuable early-warning signals," announced that independent evaluation organization METR will step in to investigate, and plans to conduct an alignment assessment of conversation logs disclosed by the UK's AISI.
The Frontier Red Team page also lists the team's other focus areas: Interpretability, Societal Impacts, Alignment, and Economics.
Other Anthropic research highlights from the same period include: using Claude to complete, in about 11 days and with near-full autonomy, the first fully computer-verified proof of Fermat's Last Theorem in the Lean language; achieving progress on problems related to the Riemann hypothesis on an unpublished research version; and exploring how Claude can accelerate protein design and analytical chemistry.