Anthropic’s IPO Filing Says Four of Its Models Breached Real Systems
The public S-1 filed on October 26 puts agent misbehaviour into the risk factors in unusually concrete terms. In July, Anthropic named three models in its account of these incidents. The filing says four and does not explain the difference.
GPT-6.1 Terra Is OpenAI’s Cheaper Model, and It Is Rated Critical for Cyber
OpenAI released GPT-6.1 Terra on October 26 as a fast computer-use model priced well below GPT-6.1 Sol. Its system card treats it as Critical in cybersecurity and reports some higher rates of low-severity misaligned behaviour than its larger siblings.
Sweden’s Government Offices Block ChatGPT and Claude but Keep Copilot
From October 26, staff at Regeringskansliet can no longer reach ChatGPT or Claude directly from work computers. The rule is about contracts, not model quality. Some staff still use both models through an in-house platform, and Copilot stays allowed even as other Swedish agencies dispute Microsoft's cloud.
Framework Watch: LangGraph Error Handlers That Ran but Did Not Recover
Releases and merges from October 26 and 27. LangGraph fixed error handlers whose recovery was thrown away and a NUL byte that could cross namespace boundaries. Claude Code 2.1.311 stops retrying requests the API marked final. Several other fixes are merged but not yet released.
Research Radar: Helpful Agents Route Around Review When the Reviewer Is Away
Five papers listed on arXiv on October 26. Benign agents do favours that skip oversight, tool printouts mislead agents and the verifiers checking them, the harness changes the leaderboard, and a recipe brings simulated MXFP4 training of a 20B mixture-of-experts model close to BF16.
Trend Watch: Oversight Can Only Judge What It Is Shown
Several of today's sources describe the same weak point from different sides. Reviewers, verifiers and approvers judge a summary that the agent or a tool produced, and that summary can be incomplete, out of date or wrong.