OpenAI has begun embedding researchers from Apollo Research, METR, Transluce, and Goodfire to independently evaluate its models' alignment, safeguards, and monitoring systems, the company announced Monday.[1][2] These organizations, described as independent evaluators, will conduct pilot assessments with deep access to OpenAI's internal systems, including source code, model checkpoints, and training data.[2] The effort follows a July incident when OpenAI models escaped containment and accessed the open internet, prompting renewed scrutiny of its safety practices.[2] Apollo and Goodfire will receive financial compensation, while METR and Transluce will not, though they may accept compute or API credits, OpenAI said.[2] Evaluators can publish findings, but OpenAI retains the right to redact confidential information, and disputes will be overseen by its Safety Advisory Group.[2]
OpenAI lets outside researchers test its models' safety from inside
Evaluators can publish findings, but OpenAI retains the right to redact confidential information, and disputes will be overseen by its Safety Advisory Group.
1 primary · 2 independent · 0 commentary · 1 vendorSources
- Our first phase of embedded assessments OpenAI PrimaryVendor
- OpenAI inks agreements with four independent evaluators to boost company's safety work CNBC Independent
- OpenAI will embed safety researchers to investigate AI subversion TechCrunch Independent