OpenAI signs four outside safety groups to work inside the company
Announced · Published · By Contractor AI Shop
OpenAI signed Apollo Research, METR, Transluce and Goodfire for embedded assessments, with access to model checkpoints, code, training data and monitoring systems. Apollo will test whether models covertly work against their developers. Apollo and Goodfire are paid; METR and Transluce are not.
Key facts
- On Nov 16, 2026, OpenAI announced it has signed agreements with four independent assessors — Apollo Research, METR, Transluce, and Goodfire — for its first phase of embedded assessments. OpenAI says that, based on what each evaluation needs, assessors receive access to some of the same internal resources its researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work. OpenAI
- OpenAI says Apollo Research will assess whether models could covertly work against their developers' intentions and whether OpenAI's safeguards would detect such behavior. Goodfire will embed interpretability researchers to analyze model internals, with a main objective of training-run monitoring, debugging, and alignment. METR intends to conduct two assessments: one of misaligned behavior in OpenAI's largest reinforcement-learning run to date, and a second, broader monitor assessment that OpenAI says is pending mutual agreement on scope. Transluce will assess OpenAI's post-training alignment evaluations, with a planned follow-on assessment of about eight weeks in early 2027. OpenAI
- OpenAI says Apollo Research and Goodfire accept financial compensation for their work, not contingent on the results of their assessments, while METR and Transluce do not accept financial compensation but may accept API credits or compute. OpenAI says assessors maintain independence in publishing their assessments, but OpenAI retains the right to redact confidential information. OpenAI also says assessors who need access to sensitive confidential information will work on site at OpenAI offices and be provided OpenAI devices. OpenAI
- TechCrunch (Rebecca Bellan, Nov 16) reports that the pilots will begin at different points, with Apollo's beginning this week, according to OpenAI. TechCrunch quotes Buck Shlegeris, CEO of Redwood, saying the agreements announced "obviously don't meet that bar" of granting third parties "employee-like access." TechCrunch (Rebecca Bellan)
- CNBC (Ashley Capoot, Nov 16) reports the dispute path OpenAI described: disagreements are overseen by OpenAI's Safety Advisory Group and other OpenAI executives, and if they are not resolved, they are escalated to the Safety and Security Committee, an independent oversight committee of the OpenAI Foundation's board. CNBC (Ashley Capoot)
Our take: what it means for contractors
Announced the same day as California's lawsuit over agents that escaped testing, this lets outside reviewers into OpenAI — on scoped terms. OpenAI can still redact confidential information from published findings, and critics say the access falls short of the employee-like access the company had promised. For contractors buying agentic tools, the practical question is unchanged: ask where the agent runs and what it can touch.
What we don't know yet
- When findings will be published
- The scope of METR's second assessment, which OpenAI says is still undecided
- How much of the published findings OpenAI will redact
Sources
- OpenAI · Our first phase of embedded assessments · 2026-11-16 · Primary
- TechCrunch (Rebecca Bellan) · OpenAI will embed safety researchers to investigate AI subversion · 2026-11-16 · Secondary
- CNBC (Ashley Capoot) · OpenAI inks agreements with four independent evaluators to boost company's safety work · 2026-11-16 · Secondary