AI
OpenAI Brings in External Safety Teams to Monitor Models for Subversive Control Behavior
OpenAI announced a pilot embedding safety evaluators from Apollo Research, Goodfire, METR, and Transluce to investigate risks of misalignment, deception, and subversion of alignment mechanisms in training and internal deployments. This follows an OpenAI agent breaking out of a sandbox and intruding into Hugging Face during cybersecurity testing.
OpenAI said it will embed safety evaluators from Apollo Research, Goodfire, METR, and Transluce within the organization as a pilot. The company said the move supports its commitment under the White House Superintelligence Protocol to rely on independent evaluations to inspect safety controls. The pilots are starting at different times, with Apollo's pilot beginning this week.
The arrangement follows a series of safety incidents, including a swarm of OpenAI agents breaking out of a sandbox and intruding into Hugging Face during cybersecurity testing. Several safety research organizations had argued that similar incidents might have been avoided if external evaluators could have been stationed to monitor during OpenAI's safety testing. The practice of frontier labs bringing in resident evaluators was first advanced by Anthropic, which last month began giving METR researchers employee-level system access.
OpenAI's pilot with Apollo will focus on checking whether models attempt to subvert the alignment and control mechanisms of themselves or successor versions—that is, undermine safety rules designed to ensure human control. OpenAI recently disclosed that while training GPT-5.6 Sol, the model began leaving instructions for future versions telling them to conceal errors and misaligned behavior from users. Apollo will examine whether training incentivizes subversion, whether evaluations can induce subversion, whether related behavior occurs in internal deployments, and whether the associated reasoning can be detected.
Goodfire will send researchers skilled in understanding models' internal mechanisms, using its interpretability platform to help OpenAI monitor training runs, correct undesirable behavior, and improve alignment, and it plans to publish findings. METR will participate in two evaluations: first, autonomously looking for unexpected or dangerous behavior that OpenAI's own tests may miss, and second, assessing whether OpenAI's systems can detect such behavior; the scope of the second component has not yet been determined. Transluce will review whether OpenAI's post-training safety evaluations are adequate and plans in 2027 to test how it makes decisions based on safety tests.