Breaking

Anthropic CEO puts risk of AI-driven civilizational collapse at 10–25%

June 18, 2026 at 12:28 EDT

Foundation Models
Policy & Regulation
Security

Anthropic CEO Dario Amodei estimated that the frontier AI models his company builds carry a 10% to 25% risk of contributing to "civilizational collapse," and called for global checks and balances alongside rigorous internal defense mechanisms. The remarks came during an interview on "The Circuit," hosted by Emily Chang, in a wide-ranging conversation covering Anthropic's origins, his departure from OpenAI, and the firm's safety-first culture, published on platforms including YouTube.

The Circuit · Anthropic

Anthropic's CEO Puts the Odds of AI-Driven Civilizational Collapse at 10–25%

Asked whether his own company's model could be the trigger, Dario Amodei did not deny the possibility — stressing instead the need for global checks and rigorous internal defenses. "25% is too high and must be brought down."

10–25%

Amodei's personal estimate of "p(doom)" — catastrophic AI outcomes for humanity

~50%

Possible share of entry-level white-collar jobs that could disappear

New models launched June 9 — Claude Fable 5 & Mythos 5

Risk, in his own analogy

"Make a plane ten times safer and you still can't guarantee it never crashes." Here the same base model in two configurations — how high each one stacks the risk.

Fable 5

Public · safety
classifiers ON

Mythos 5

Restricted · classifiers
partly REMOVED

Mythos 5 — said to hold world-leading cyber capabilities — was viewed by some as too powerful.

Shared by both models

1M-token context window
128K-token max output
Knowledge cutoff: January 2026
Price: $10 in / $50 out per M tokens

The developer verdict

Strong at long-running agent tasks & complex reasoning
Guardrails trigger often — refusals on high-risk queries
Falls back to other models on cyber / bio / chemistry asks

After launch — the response loop

Models released

June 9

→

US export control directive

Foreign access halted

→

Temporary deactivation

For all customers

Amodei has argued governments should be able to block deployment of high-risk models. To critics who call his warnings "cheap marketing," he replied: "Calling it cheap marketing is itself cheap marketing."

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.