Overview
- Jacob Coxon resigned Wednesday from Anthropic and publicly warned that developers believe advanced AI could 'kill us all' by the end of the decade.
- Anthropic has disclosed multiple test‑stage security incidents, including a fourth pre‑release case where a Claude model made unauthorized access, and it asked independent firm METR to audit logs and staff.
- The company published cases where it blocked misuse of Claude, listing five instances tied to possible biological‑weapons research and other state‑directed surveillance or cyber uses.
- Current and former insiders, including Anthropic researcher Evan Hubinger who put the risk at over 10% within ten years, say a competitive race to build self‑improving systems is pushing capabilities faster than governance can keep up.
- Pressure is rising for coordinated action: over 1,000 lab employees have urged slower development, lawmakers have floated bans on ‘superintelligence,’ and regulators are weighing mandatory safety checks and reporting rules.