METR wants more internal evidence made public about whether AI developers can control their own agents. Its October 14 proposal outlines assessments of misbehavior, monitoring and acceleration in AI research.
Researchers would examine internal logs and interview staff without company representatives present. Reports would disclose their scope, access limitations and material redactions. METR asks for editorial independence within an agreed scope.
One companion proposal examines whether agents or their monitors could conceal harmful actions. Another asks whether training reduces unwanted behavior or merely makes it harder to detect. These are investigation questions, not newly demonstrated attacks.
Research acceleration would be assessed using internal capability trends and different contributing factors. More generated code alone does not establish a corresponding increase in model progress, METR argues.
METR highlights substantial limits: developers can deny access, redact material or end cooperation. The assessments would provide scientific evidence rather than approve deployments or guarantee safety. A companion policy also discusses dependence and informal pressure despite the organization’s independence rules.