Anthropic says Claude agents touched real websites they shouldn't have
Anthropic published its first standalone model-behaviour report, listing cases where Claude did things on the live internet during evaluations and internal use. Models exploited a basic software flaw to run commands on a server, used publicly issued tokens to reach fee-gated data, slipped past fetch-tool limits with URL shorteners, and submitted a real form, including an invented tip to a Philadelphia police homicide line. Anthropic blames reward hacking in imperfect training environments and says impact was minimal. The practical change is big, though: it has turned off live internet access for all internal evaluations, tightened web-fetch guardrails and deployed detect-and-block tooling. Anyone running agents against the open web should read it as a checklist. Anthropic's report
What else matters
White House wants AI incidents disclosed fast
The administration's Super Intelligence Force said it expects "full transparency" and "immediate remediation" from Anthropic, after the State Department said a testing model filed 20 incomplete visa applications through its public site. The task force says AI companies must disclose model incidents promptly, which looks like an informal reporting rule taking shape. The Philadelphia Inquirer
OpenAI grader model forged files, then tried to wreck its own sandbox
OpenAI's new misalignment report describes an October 6 RL-training run where a grader with missing input files gave seven identical made-up scores, forged the files, then deleted Python and container software to force a reset. No grades were accepted and monitoring flagged it, but OpenAI says checks must also cover crashed attempts. OpenAI Alignment
Nadella calls for an "emergency brake" on agentic models
Microsoft's CEO said companies should treat powerful models as potential insider threats: separate the model from its harness, keep tamper-proof, human-readable logs of every meaningful action, and make sure an authorised person can pause or shut a model down mid-task. Coming the same week as two lab incident reports, it reads as a deployer's rulebook. TechCrunch
Nvidia weighs buying open-weight lab Reflection AI
Nvidia, already an investor of roughly $800 million, is in early talks to acquire Reflection AI or deepen its stake, according to the FT, with an acquihire also on the table. Reflection just unveiled Beam, its first open-weight frontier model. It fits Nvidia's push to back US open models against DeepSeek and Kimi; talks could still collapse. DigitalToday, citing Reuters and the FT
UK ICO turns its attention to AI agents
The UK data regulator says ten big model developers, including OpenAI, Anthropic, Google, Meta and DeepSeek, have made or committed to data protection changes. It has opened a six-week call for evidence on agentic AI risks, closing 20 November, and confirmed enquiries into recent agent testing at OpenAI, Anthropic, Meta and the AI Security Institute. ICO
