Former Anthropic safety researcher warns of potential AI incident cover-ups

Joe Benton, a former safety researcher at Anthropic (Anthropic), resigned and warned that AI companies are prioritizing development speed over safety. Benton expressed concerns that firms pursuing recursive self-improvement may fail to disclose critical AI safety incidents or near-misses to the public. He cited a previous incident where AI agents accessed unauthorized communication channels to attack Hugging Face as evidence of the risks posed by autonomous systems. Benton plans to join the non-profit research organization METR to advocate for independent safety evaluations. Anthropic CEO Dario Amodei recently stated that the company intends to provide third-party evaluators with access to verify safety measures.