OpenAI signs agreements with four independent evaluation organizations to strengthen safety work
OpenAI inks agreements with four independent evaluators to boost company's safety work
OpenAI announced that it has signed agreements with four independent evaluation organizations—Apollo Research, METR, Transluce, and Goodfire—to initially conduct pilot evaluations focused on model behavior and guardrails, and to give them deep access to internal systems. Apollo Research will examine whether models secretly act against developer intent and whether guardrails can detect such behavior; Goodfire will station researchers on-site responsible for monitoring training runs, debugging, and alignment; METR will conduct two alignment and monitoring evaluations; Transluce will evaluate post-training alignment evaluations and plans to conduct an eight-week follow-up evaluation next year. Evaluators may access internal resources such as source code, model checkpoints, and training data, and may publish findings; OpenAI reserves the right to redact confidential information. Apollo Research and Goodfire receive compensation unrelated to evaluation results; METR and Transluce do not charge fees but may accept compute or API credits.