Anthropic safety team member announces departure, urges more AI transparency

On September 11, a former member of Anthropic's safety team said they had left two weeks earlier and planned to join METR for independent AI risk evaluations. They warned that a company could lose control of its systems without the public knowing. Their proposed guardrails include disclosing progress toward recursive self-improvement—AI systems improving themselves—reporting safety incidents and near-misses, and meeting minimum safety standards with independent guarantees that those standards are met.