UNCOS

Search Uncos

OpenAI will embed safety researchers to investigate AI subversion

OpenAI will embed safety researchers to investigate AI subversion

image via TechCrunch

November 16, 2026, 5:02 PM

  • •Apollo Research will assess whether OpenAI models attempt to subvert their own or successor models’ alignment and control mechanisms.
  • •Goodfire will use interpretability tools to help monitor training runs, identify unwanted behavior, and improve alignment.
  • •METR will run two evaluations: one for dangerous or unintended model behavior and another on whether OpenAI’s own systems can detect it.
  • •Transluce will examine whether OpenAI’s post-training safety evaluations are sufficient and plans a 2027 test of release-decision processes.
  • •The article says OpenAI is limiting evaluator access to agreed claims, models, staff, and time periods, while reserving some rights to withhold information for security, feasibility, or IP reasons.
  • •Critics including Redwood CEO Buck Shlegeris argue the arrangements do not yet amount to true employee-like access for independent evaluators.

OpenAI said it will embed outside evaluators from Apollo Research, Goodfire, METR, and Transluce to examine whether its AI systems show dangerous or deceptive behavior. Apollo’s pilot will focus on whether models try to subvert alignment and control mechanisms, while the other groups will study interpretability, unintended behavior, and the adequacy of OpenAI’s safety testing. The article says these arrangements follow recent safety incidents and growing pressure for independent assessments during training and deployment, not just before release. It also notes that the pilots stop short of full employee-level access and give OpenAI some control over information access and public disclosure, raising concerns about evaluator independence.

Read original article

Entities Mentioned

Rebecca BellanJohn SteidleyBuck Shlegeris

Topics Covered

AIai safety researchOpenAIArtificial IntelligenceAI SafetyModel EvaluationInterpretability

Comments (0)

No comments yet.