Our first phase of embedded assessments
- ID
- 47084
- Status
- summarized
- Published
- 16 Nov 2026, 4:00 PM
- Fetched
- 17 Nov 2026, 1:43 AM
- Provider
- OpenAI News
- Category
- ai-labs
- Original URL
- https://openai.com/index/our-first-phase-of-embedded-assessments
- Source URL
- https://openai.com/news/rss.xml
Summary
- Score
- 4.5
- Created
- 17 Nov 2026, 1:43 AM
- Tags
- Audience
- ai_ml_learnersai_agent_userssaas_founders
What happened
OpenAI has signed agreements with four independent assessors — Apollo Research, METR, Transluce, and Goodfire — for a first phase of "embedded assessments," giving them access to internal resources including model checkpoints, source code, training and evaluation data, monitoring systems, and discussions with the researchers and engineers responsible for the work. The initial scope is alignment and safeguards for models during training and internal use, complementing OpenAI's safety cases, and supports commitments under the White House Accord on Super Intelligence. Apollo Research's pilot will test four claims in two stages, starting with whether training incentivized subversion and whether strong evaluations and red teaming elicit subversion.
Why it matters
The stated scope is alignment and safeguards during training and internal use — not the deployed API behaviour your app depends on. So if your risk register or customer-facing security docs lean on "the lab says the model is safe," this announcement does not cover your agent's runtime actions, your prompts, or your data flows; you still need your own evals and guardrails. Treat it as governance context to cite, not a control you can inherit.
Discussion angle
Apollo's four claims are checkable in principle, but OpenAI chooses the assessors and grants the access. Ask the room: which of the four subversion claims would you actually trust from this setup, and what would an assessor need that isn't in the listed resources (checkpoints, source code, training data, monitoring systems) to falsify them?