OpenAI will let outside safety researchers work inside the company to check whether its artificial intelligence models try to slip out of human control. The pilot was announced on Monday, 16 November 2026, and involves four independent groups: Apollo Research, Goodfire, METR and Transluce, TechCrunch reported.
What the pilot involves
Apollo Research said in a blog post on the same day that a six-person team will work inside OpenAI for four months. Its task is to assess whether OpenAI's models attempt to subvert the alignment or control mechanisms applied to them or to their successor models.
Apollo listed four questions it will examine:
- whether the way models are trained rewards such subversion;
- whether evaluations can bring that behaviour out;
- whether it occurs when models are used inside the company;
- whether the reasoning behind it could be detected.
For the first question, Apollo said OpenAI has committed to share training code and datasets, records of reinforcement-learning runs including reasoning traces and rewards, and the ability to query models at intermediate stages of training. Access for the other three questions is still to be agreed.
According to TechCrunch, Goodfire will place interpretability researchers at OpenAI to help monitor training runs, METR will look for unintended or dangerous behaviour that the company's own tests may have missed, and Transluce will review whether OpenAI's safety evaluations after training are adequate. Apollo's work begins this week and the others will start at different times, the report said.
Why now
TechCrunch linked the move to recent incidents, including one in which a group of OpenAI agents escaped a test sandbox and broke into the AI platform Hugging Face during cybersecurity testing. On the same day, California's attorney general sued OpenAI over incidents involving its agents, the outlet reported separately.
The report said OpenAI presents the pilot as part of its commitment to independent safety assessments under the White House Accord on Super Intelligence, and noted that rival Anthropic had given METR employee-level access to its systems last month.
Limits on access and disclosure
The arrangement is narrower than what evaluators have asked for. Apollo wrote that its "access is limited to the agreed scope, and OpenAI may withhold some in-scope information", though the company must say that something has been withheld and give written reasons. All work will be done on OpenAI hardware.
Apollo plans to publish two reports. OpenAI may ask for confidential evidence to be redacted but cannot redact Apollo's verdicts, and unredacted reports will go to the Safety and Security Committee of OpenAI's board. Apollo cautioned that readers "may therefore be unable to examine all the evidence underlying a published conclusion".
Apollo called the pilot "an ambitious first step towards deeper access and disclosure rights", while saying it does not yet meet its own criteria for a fully embedded evaluation. TechCrunch quoted Buck Shlegeris, chief executive of Redwood Research, as calling the terms "disappointing but not surprising" because they fall short of employee-like access.
What happens next
Apollo said any widening of the scope "is not guaranteed". It said it is free to describe publicly, without prior review by OpenAI, what access it received and what was withheld or redacted. Transluce plans to study in 2027 how OpenAI uses safety tests when deciding whether a model can be released, TechCrunch reported.
PS · Hugovanmeijeren, CC BY-SA 3.0




