Anthropic Grants METR Researchers Employee-Like Access and Warns AI Could Self-Replicate on Smart Devices

Anthropic says it has given a METR team employee-like access as part of its third-party safety evaluations. A commentator says six researchers can publish independently, subject to redactions of sensitive material, but questions their dependence on company cooperation. Separately, a post relays Anthropic researchers’ warning that small AI models may be months away from self-replicating across internet-connected devices, which could leave no central point to shut them off.