OpenAI launches embedded assessments with external labs
Reported by OpenAI, Blockchain News, FourWeekMBA and 3 more outlets6 sources ↓
Latest in this story: OpenAI plans to embed Apollo researchers to investigate whether models undermine safety measures
OpenAI says Apollo Research, METR, Transluce and Goodfire will assess its models and safety systems, with access to internal resources that can include unreleased models, code, training data and monitoring systems.
Key points
- Four organizations will conduct separate assessments of OpenAI’s models and safety systems
- Access can include unreleased models, checkpoints, source code, training data and monitoring systems
- Assessors can publish findings, while OpenAI may redact confidential information or delay reports to address risks
- OpenAI and Transluce plan an approximately eight-week follow-up assessment in early 2027
Sources
- OpenAIRead ↗Our first phase of embedded assessments
- Blockchain NewsRead ↗OpenAI launches embedded assessments with external labs
- FourWeekMBARead ↗OpenAI Lets 4 Safety Labs Inside Its Models; Apollo Embeds 6
- CNBCRead ↗OpenAI inks agreements with four independent evaluators to boost company's safety work
- Unite.AIRead ↗OpenAI Signs Four Independent Assessors for Embedded Safety Reviews
- The DecoderRead ↗OpenAI opens its AI systems to four independent on-site auditors for the first time
Details
The assessments have different initial scopes. METR will examine a reinforcement-learning training run that OpenAI describes as its largest to date; Goodfire will study model internals and training-run monitoring; and Transluce will review post-training alignment evaluations. The Decoder reports that assessors may work on-site and speak confidentially with developers.
The organizations can publish their findings, subject to OpenAI’s right to redact confidential material and request delays to address safety or security risks. An independent board committee will receive unredacted reports and oversee the process, with disputes escalated to it if they remain unresolved. OpenAI and Transluce also intend to conduct a broader, approximately eight-week follow-up review in early 2027.
Written from coverage as of
How this story developed
- OpenAI signs agreements with four outside labs for embedded safety assessmentsthis story
- OpenAI plans to embed Apollo researchers to investigate whether models undermine safety measures
Images
- CNBC
- Unite.AI
- Blockchain News
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh