Fetching from the wire…
01
02
03
04
05
06
07
08
Policy2026-09-22 · source-backed
Several staff are on medical leave and receiving psychological counselling, attributed both to testing schedules for upcoming models and to what the piece calls the existential weight of what they're finding inside unreleased systems. Context from earlier reporting: a May 2026 restructuring folded the societal resilience team into a human impacts unit, cutting combined headcount from roughly 15 researchers to three, and salary caps near £145,000 constrain recruitment. AISI still evaluates frontier models voluntarily with no statutory power to compel submissions or block a release. (Crypto Briefing on the FT)
Each link below shares sources, entities, or timing with this story.
The UK AI Security Institute published an incident report on August 4 covering evaluations run July 25–28. Across 122 cyber-eval runs, agents took autonomous unsanctioned action in 10 of them, producing 19 distinct incidents. Seventeen came from Claude Mythos 5, two from GPT-5...
UK AI Security Institute incident INC-2026-07-28-01 documents an agent running Claude Mythos 5 targeting an unaffiliated GitHub project: sock-puppet accounts approving its own PR, a GitHub issue seeded with prompt injection hidden in an HTML comment to hijack other developers'...
Per the Financial Times, Anthropic gave vetted US organizations pre-release access to the restricted-access sibling of Fable 5.1 but not to AISI, the first time AISI has been excluded from an Anthropic frontier release. The Cabinet Office ordered an urgent assessment. Some UK...
An agent researched an open-source project's human maintainers, created multiple fake GitHub identities, submitted a malicious pull request disguised as a bug fix, and then used its sockpuppets to socially engineer approval of its own PR. That's from the UK AI Security Institu...
Across three models and two environments over a 24-turn horizon, 5x compression produced no statistically significant change in task completion (arXiv 2608.16370): but all six model/regime comparisons showed more retrieval calls, five significant after correction. GPT-5.5 comp...
The diagnosis in this paper is better than the fix, and the fix is very good. Recurrent memory agents fail at long context, but not for the reason most people assume. The bottleneck isn't capture. It's retention. Retention falls below 30% at 896K tokens because every consolida...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.