Anthropic safety chief quits, warning ‘world is in peril’ from AI risks

Sharma’s resignation is a rare public flashpoint inside a top frontier lab, and it lands at a moment when Anthropic is raising tens of billions of dollars and positioning itself as a safety‑first alternative to OpenAI. His letter doesn’t allege specific misconduct, but it does suggest that even organizations built around AI safety struggle to align day‑to‑day decisions with their stated values once capital, competition and timelines tighten. For Race to AGI readers, this is a live case study in how governance and culture hold up under extreme scaling pressure.

The move also heightens scrutiny of Anthropic’s safeguards work just as it leans hard into agentic systems like Claude Opus 4.6 and Claude Code, tools explicitly designed to automate more complex workflows. If the head of safeguards feels the world is in peril and chooses to step away rather than keep shaping those systems from the inside, that will amplify calls for external oversight and more transparent safety cases around powerful models. Competitively, this doesn’t immediately change Anthropic’s technical trajectory, but it may influence how regulators, partners and talent perceive its ability to self‑regulate while racing toward AGI‑class capabilities.