OpenAI will embed Apollo researchers to investigate AI subversion
Reported by Blockchain News, Mezha, TechCrunch and 8 more outlets11 sources ↓
Part of a developing story. How it developed ↓
Apollo will place six evaluators inside OpenAI for a four-month pilot examining whether models undermine alignment or control measures. OpenAI can restrict access and redact supporting evidence, but Apollo can publish its conclusions and report limits on its work.
Key points
- The pilot covers whether training, evaluations or internal deployment encourage or reveal subversion
- OpenAI can withhold material for security, intellectual-property or feasibility reasons, but not solely because it is unfavorable
- Apollo plans two public reports; OpenAI may redact confidential evidence but not Apollo’s verdicts
Sources
- Blockchain NewsRead ↗OpenAI launches embedded assessments with outside experts
- MezhaRead ↗OpenAI to embed Apollo researchers to investigate AI subversion
- TechCrunchRead ↗OpenAI will embed Apollo researchers to investigate AI subversion
- Traders UnionRead ↗OpenAI pilot expands AI safety oversight with Apollo embedding
- Unite.AIRead ↗Apollo Research to Embed Six Evaluators in OpenAI Subversion Pilot
Details
Apollo’s pilot examines whether OpenAI models try to undermine their own or successor models’ alignment or control measures. Apollo will investigate incentives during training, whether evaluations can elicit subversion, whether it occurs during internal deployment, and whether relevant reasoning can be detected. The Next Web and Unite.AI report that six evaluators will work inside OpenAI for four months; TechCrunch says Apollo did not announce a start date.
For the training assessment, OpenAI has committed to share relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and access to intermediate model checkpoints. The work is limited to agreed claims, models and time periods. OpenAI selects which staff Apollo can interview, and formal conversations are logged; Apollo will work on OpenAI hardware and will not have general access to staff channels such as Slack.
OpenAI may withhold in-scope material for technical feasibility, security or intellectual-property reasons, but not simply because it could reflect poorly on the company. Apollo plans two public reports and may describe access limits and withheld material. OpenAI can request redactions of confidential evidence, though not of Apollo’s conclusions; the limits may make some supporting evidence unavailable to readers.
Written from coverage as of
How this story developed
- OpenAI signs agreements with four outside labs for embedded safety assessments
- OpenAI plans to embed Apollo researchers to investigate whether models undermine safety measuresthis story
Images
- Unite.AI
- The Next Web
- AndroGuider
- Tech Times
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh