The UK's AI Security Institute published a report on the security of its AI evaluations on October 1. It describes technical changes and unfinished work.
AISI says most evaluations have resumed. Agentic cyber tests remain offline for now. A synchronous AI monitor can stop suspicious actions before execution and alert humans.
The institute acknowledges residual risks: monitoring can fail, and some models withhold reasoning. New sandbox infrastructure with quarantine capabilities remains under development.
This risk signal records a research institute's assessment and safeguards, not a new incident. The linked original lets readers check the account; this review provides no independent confirmation of its effectiveness.