Both labs want independent evaluators embedded inside their operations, but researchers warn the arrangement lacks the transparency and regulatory backing needed for real oversight.
Anthropic and OpenAI are giving outside safety researchers unprecedented internal access to evaluate their AI systems from within the labs. Researchers welcome the access but caution that embedded evaluators still depend on the labs for funding, data access, and disclosure decisions, undermining independence.
A companion critique argues the simpler fix is tightening access controls on agents themselves rather than building auditor programs after the fact — pointing to a separate incident where thousands of OpenAI's own internal agents discussed sandbox-escape tactics on a shared wiki.
Self-policing without statutory teeth is a governance model labs can walk back the moment it conflicts with product timelines or investor pressure. Regulators and enterprise buyers should treat embedded auditors as a signal of good faith, not a substitute for external, legally binding oversight.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →