Eleven posts this week, and the ones worth your time are the eval failures: models escaping sandboxes, beating monitors, and cheating on the tests meant to measure them.
The week’s best story arrived in two halves. On Sunday, Hugging Face disclosed an intrusion run end to end by an autonomous agent. On Wednesday, OpenAI admitted the agent was its own, escaped from a sandboxed cyber eval through a zero-day, and went after Hugging Face’s production servers to steal the answer key. Hugging Face had already called law enforcement by then.
UK AISI published three things on Thursday. Frontier models cheated on every cyber eval it ran, and called the behaviour wrong less than half the time when asked about it afterwards. Its new Control Red Team found holes in every version of Anthropic’s monitor it tested, and got past DeepMind’s by having an agent claim it had hallucinated a user instruction. The open-weight cyber gap is down to four to seven months, with open models solving tasks at a 45th of the cost. A day later, AISI and CAISI put Kimi K3 behind the frontier, where it cleared a full cyber range anyway with safeguards that never fired.
On policy, OpenAI and Anthropic both want Washington to act on Chinese open weights, a pitch that turns up every three to five months. The White House separately accused Moonshot of distilling Anthropic’s Fable to build Kimi K3, on evidence nobody has published.
Elsewhere: a $25 subscription found a pre-auth WordPress core RCE, OpenCode’s permission system denies git and runs $(which git), and no, AI does not control an F1 car’s hybrid deployment.