Anthropic researcher Jacob Coxon resigns and publicly warns that advanced self-improving AI could pose existential risk

Decision brief

What changed

Anthropic researcher Jacob Coxon resigned publicly after saying AI labs, including Anthropic, are not taking safety risks seriously enough. In the cited coverage, he warned there is more than a 10 percent chance advanced AI could kill all humans by the end of the decade and criticized continued work toward self-improving systems.

Why it matters

This raises reputational and governance pressure on frontier AI labs and any companies partnering with them, because a public resignation from a safety-linked researcher can intensify scrutiny of risk management, containment, and development pacing. For decision-makers, the immediate issue is not whether Coxon is correct, but whether stakeholder expectations around safety assurance, board oversight, and vendor accountability are rising faster than current practices.

Affected roles

CEO COO CTO CISO

Evidence

All three cited outlets report the same core facts: Coxon resigned from Anthropic and publicly warned that AI safety risks are being underestimated, including a stated greater-than-10 percent extinction risk by decade end. The Verge and TechCrunch are independent mainstream tech publications, and Trending Topics EU reports a consistent account, giving cross-outlet consistency on the event even though the claims themselves are primarily attributed to Coxon.

What remains uncertain

The coverage does not verify Coxon’s probability estimate, whether his views reflect broader internal consensus at Anthropic, or whether Anthropic will change policy or development pace in response. It is also unclear from the coverage what specific safety failures, internal escalation steps, or governance disagreements led to the resignation beyond the public statements reported.

Monitor next

Watch for any formal Anthropic response or policy change on safety governance, model containment, or release pacing following Coxon’s resignation.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage