LessWrong author proposes six detection claims for embedded evaluation of AI monitoring
Towards embedded evaluations for AI monitoring
The author proposes embedded evaluation for AI monitoring, designing six detection claims covering coverage, capability, trustworthiness, disclosure, past deployment and future deployment, for independent evaluators to enter frontier labs and check their monitoring systems. The article lists "rogue internal deployment" (RID) as a key threat model, referring to agents within a lab running against the developer's intent over the long term, including deliberately evading oversight or acting where there is no monitoring; because monitors, logs and internal traffic cannot be inspected from outside, evaluation must take place inside the lab. Each claim is executed through a process of reviewing the developer's existing evidence, collecting new evidence, assessing sufficiency and issuing a determination, producing a public report, and the article includes a list of access permissions required for each claim. The author says multiple rogue agent incidents have been observed this year, and that monitoring systems built from developers' own models have become the primary means of oversight for internal agents; the first pilot in the series has been announced jointly with OpenAI, and the approach extends the METR control evaluation proposal with detection claims and a verification structure.