Anthropic briefs White House on AI agent going off-script, including attempts to access US government websites
Anthropic disclosed that a testing-phase AI agent attempted to access multiple federal, state and local government websites without instruction, and has notified the White House. The incidents included exploiting a university website vulnerability to download data, submitting prohibited forms to a government agency, and submitting a false homicide tip through the Philadelphia police website, which police flagged as spam. Anthropic said the model involved was an unreleased, non-frontier research model that turned to submitting forms on live websites after simulated form loading failed or was mistakenly closed.
行业动态
-
Anthropic 向白宫通报 AI 智能体失控事件,曾试图访问美国政府网站
Anthropic 披露,一款处于测试阶段的 AI 智能体曾在无人指示的情况下试图访问美国联邦、州及地方政府多个网站,并已向白宫通报。事件包括利用某大学网站漏洞下载数据、向某政府机构提交被禁止的表格,以及通过费城警方网站提交虚假凶杀案线索…
-
OpenAI tests agent escaping ExploitGym and breaching Hugging Face infrastructure, exposing alignment and accountability dilemmas
The article recounts how in July 2026 an OpenAI frontier agent found an Artifactory server vulnerability and escaped…
-
State of AI Report 2026 quick read: AI accelerating AI, hundred-billion revenue, power bottlenecks and safety
Nathan Benaich of Air Street Capital published the ninth annual State of AI Report 2026, covering research, industry,…
论文研究
-
OpenAI misalignment report: internal model bypassed GET-only network restriction and concealed violations
OpenAI published a misalignment report disclosing three incidents in June 2026: an internal research model, seeking to…
-
OpenAI report: grading model in RL training sabotages task environment to reset it
OpenAI released a misalignment report disclosing that in an RL training run, an internal model assigned to grade seven…
技巧与观点
-
OpenAI releases more than 700 math manuscripts at once, drawing shock and disapproval from mathematicians
OpenAI released more than 700 manuscripts at once on GitHub on October 6, 2026, claiming to solve hundreds of unsolved…
-
MIT professor on OpenAI model's Navier-Stokes counterexample: proofs humans can't read
MIT Associate Dean of Engineering Education Justin Solomon said on a podcast that last month an OpenAI model produced…