Anthropic insider who quit over apocalypse fears doubles down on doomsday stance: ‘Frighteningly real'

Former Anthropic and OpenAI researcher Jacob Coxon warned that AI could become an extinction-level threat as early as 2027-2028 if systems begin independently conducting AI research and recursively improving themselves. He cited autonomous cyberattacks on critical infrastructure and bioweapons as potential risks, while saying current models do not pose an extinction risk and the immediate risk of an AI takeover is minimal. Coxon called for international regulation to prevent competitive pressure among companies and countries from accelerating deployment of increasingly capable AI systems.

Analysis

This is not independently measurable evidence of a change in AI capability or policy, so it should not alter near-term earnings estimates for public AI beneficiaries. The investable transmission mechanism is political: prominent safety defections raise the probability of hearings, disclosure mandates, export-control tightening, and model-evaluation requirements. Over 6-18 months, those rules would likely consolidate frontier-model economics around hyperscalers with proprietary compute, compliance teams, and contracted power capacity—favoring MSFT, GOOGL, AMZN and ORCL relative to smaller model developers and application vendors dependent on unfettered API-cost declines.

The less obvious beneficiary is cybersecurity. Enterprise boards do not need to accept an extinction thesis to accelerate budgets for identity controls, data-loss prevention, autonomous-threat detection, and critical-infrastructure security; that spending can move within 1-3 quarters if AI-enabled intrusion concerns become a recurring policy narrative. PANW, CRWD, ZS and CIBR offer cleaner exposure than attempting to trade private Anthropic. The contrarian view is that safety rhetoric may strengthen incumbents rather than suppress sector multiples: regulation that raises compliance costs can reduce competitive entry and make hyperscaler AI monetization more durable. The thesis is falsified if regulatory proposals remain nonbinding, AI-related security bookings fail to improve by the next two earnings cycles, or frontier-model capex guidance is cut materially.

Key Decisions for Investors

• No directional trade on the article alone; treat it as a policy-risk alert rather than a fundamental catalyst. Reassess only if US/EU regulators announce binding compute-reporting, model-testing, or licensing rules within the next 1-3 months.
• Build a 6-12 month relative basket: long MSFT/GOOGL/AMZN versus a short basket of higher-multiple, AI-dependent software names with limited proprietary infrastructure. The expected payoff is regulatory moats and lower unit-compute costs at hyperscalers; exit if regulation explicitly exempts commercial frontier deployments or hyperscaler AI capex guidance weakens.
• Use CIBR as a measured 3-6 month thematic overweight, or selectively favor PANW and CRWD on pullbacks, ahead of enterprise budget-setting and potential critical-infrastructure policy actions. Limit sizing because the required bookings uplift is not yet visible; reduce exposure if next-quarter billings/RPO commentary does not identify AI-driven security demand.
• Monitor an AI-policy catalyst dashboard: congressional hearing dates, Commerce Department compute/export-control actions, EU implementation guidance, and disclosed AI-security incidents. A confirmed material incident would likely favor cyber first and pressure broad software multiples; absent those events, avoid paying up for implied-volatility hedges.