Two Anthropic insiders say advanced AI could “kill us all” within ten years—and their own company’s rulebook treats catastrophe as a real risk, not a movie plot.
Story Snapshot
- Named Anthropic researchers warned of a double-digit chance AI wipes out humanity this decade.
- Anthropic’s safety policy openly targets “catastrophic risks” and tightens controls as models grow.
- Company disclosures show testing for bio, cyber, and autonomous danger before model release.
- Skeptics argue current systems are not close to extinction-grade capability yet.
Insiders put a number on extinction risk
Evan Hubinger, an Anthropic alignment scientist, wrote that he “earnestly” believes advanced AI could kill all humans, putting the risk above one in ten within the next decade. Former Anthropic and OpenAI researcher Jacob Coxon told television audiences that leading builders fear AI could “kill us all by the end of the decade”. These are not anonymous whispers. These are named researchers with direct experience at frontier labs, stepping forward with clear numbers and plain English.
The bluntness matters. Much AI debate hides behind foggy words. Here, the claims are crisp and near-term. The number is not a published measurement with a method section. It is a judgment call. Yet it lands with force because the messengers build the tools. They work on alignment, the core problem of making powerful systems follow human goals. When the people who pull the fire alarm also hold the blueprints, it moves markets and policy in a hurry.
Anthropic’s own policies treat catastrophe as a live risk
Anthropic’s Responsible Scaling Policy spells out a ladder of “AI Safety Levels” that ratchet up requirements as capability rises. The document targets “catastrophic risks” from advanced systems and binds deployment to stricter safety, security, and operational controls at each step. This is not a press release flourish. It is a rules-of-the-road plan that says, in effect, “as models grow teeth, we add thicker cages and stronger locks” to limit worst-case harm.
Anthropic’s Transparency Hub describes pre-release tests for dangerous capabilities in chemical, biological, radiological, nuclear, cybersecurity, and autonomous behaviors. That means the lab checks for the exact doors that a runaway or misused system would try to kick open. A company does not spend money building those brakes unless it sees a hill ahead. The existence of the framework does not prove doom is likely. It proves the lab treats the tail risk as a governance problem that must be managed now.
Competing view: today’s models are not doomsday machines
Many credible observers push back. Coverage by the Canadian Broadcasting Corporation reported that most experts do not think current systems are advanced enough to pose an existential threat, or even close to it. That stance fits common sense: chatbots that write emails and make coding errors are not Skynet. The counterpoint urges focus on concrete harms we can see—fraud, deepfakes, hacks—without jumping to extinction claims before the evidence arrives.
That caution is healthy. The strongest insider quotes are personal forecasts, not lab-certified odds. The record here does not include technical proof of self-improving runaway behavior. The hacking and misuse examples in the news do not bridge, by themselves, to extinction. Prudence says keep the bar high for extraordinary claims, keep testing, and do not regulate by panic. But prudence also says lock the doors before the storm, not after.
What aligns with conservative instincts: set guardrails, demand proof, refuse naïveté
Policy should match three simple American conservative values: responsibility, accountability, and deterrence. Responsibility means developers earn the right to scale by passing tough, independent safety tests tied to real failure modes. Anthropic’s scaling policy is a start; it should be verified by outside auditors, not trusted on faith. Accountability means clear lines: if a company ships a tool that helps build a bio threat or enables autonomous cyber attacks, it faces real costs.
Questioning whether someone’s behavior is consistent with his stated risk assessment is not dodging the object-level question.
It is part of it.
If Dario says frontier AI could plausibly cause catastrophic or existential harm, then asking what Anthropic itself should do…
— Cautious Optimism (@CautiousOptimi4) September 15, 2026
Deterrence means preparation for hostile use by bad actors and state rivals. The tests Anthropic lists—bio, cyber, autonomy—map to threats our adversaries would exploit first. That is where Congress and agencies should drill down. Require pre-deployment evaluations, red-team reports, and incident logs the public can verify without exposing sensitive details. Make continued scaling contingent on passing those gates. This is how we treated aviation and energy. It is how we should treat frontier AI.
The bottom line: act like adults while the clock runs
Two facts can live together. First, the extinction number from insiders is not a proven statistic. Second, the people closest to the engine are the ones waving us off the tracks. When a lab writes “catastrophic risk” into its operating manual and ties releases to harder safety tests, that is a visible signal to take the downside seriously today. Build strong brakes. Verify them. Then earn the right to go faster. That is not fear. That is discipline.
Sources:
youtube.com, bbc.com, www-cdn.anthropic.com, cbc.ca, anthropic.com, finance.yahoo.com
