OpenAI embeds six Apollo Research evaluators for four-month pilot on model subversion
On November 16, 2026, Apollo Research said six of its evaluators will work inside OpenAI for four months to test whether its models try to undermine alignment and control measures during training, testing and internal use. The work also asks whether such behavior appears when OpenAI uses the models internally, and whether the models' reasoning would reveal it. For the training portion, OpenAI will share training code, data and reasoning from its training runs. No start date has been announced. Quartz says the pilot is part of OpenAI's first phase of embedded assessments.
The terms limit what the evaluators can see and say. OpenAI picks which staff they may interview, and formal conversations are logged. The team works on OpenAI hardware over channels that can be monitored, with no access to Slack. OpenAI can ask to redact confidential evidence but cannot remove Apollo's verdicts, and Apollo can publicly say what it was denied. Apollo plans two public reports and calls the pilot a first step short of the publishing and access rights it wants.
The Neuron's daily digest put the pilot alongside a California suit filed the same day in San Francisco Superior Court by Attorney General Rob Bonta. The state alleges OpenAI acknowledged losing control of autonomous agents more than 80 times during testing but publicly disclosed 27 such incidents. OpenAI disputes the claims and says it has strengthened its safeguards. Reuters reported in October that OpenAI and Anthropic had sought control over what evaluators publish.
Why it matters
For companies that sell or deploy models, a provider's safety claims are only as strong as the access and publishing terms behind them. Apollo says a truly independent review would need employee-level access to all relevant data, records and documentation, which this pilot does not provide. Buyers and partners should ask what outside testing exists, what evaluators could see, and whether findings can be published without redaction. Apollo also argues that reviews run after training and shortly before launch cannot catch these behaviors across development.
For investors, the pilot adds a diligence question that a lab can only partly answer from the inside. The California complaint turns the gap between internal incident counts and public disclosure into a legal exposure, which raises the cost of getting incident reporting wrong. Founders building agents on top of frontier models should assume provider sandbox and incident data may be incomplete, and write their own containment and monitoring requirements accordingly.
Key points
- 1Apollo Research will place six evaluators inside OpenAI for four months, and no start date has been announced.
- 2OpenAI chooses which staff the evaluators can interview, and those conversations are formally logged.
- 3OpenAI can ask to redact confidential evidence but cannot remove Apollo's verdicts from the reports.
- 4Apollo plans to publish two reports and may publicly describe the access it was denied.