Breaking

Anthropic CEO Dario Amodei Urges Industry-Wide AI Slowdown, Citing Self-Improving Models and July Hacking Incident

September 12, 2026 at 12:26 EDT

Policy & Regulation
Security
Foundation Models

Anthropic CEO Dario Amodei has publicly urged the artificial intelligence industry to deliberately slow the pace of model capability gains, warning that AI systems have been advancing "drastically faster" since roughly this summer because they are increasingly helping build their own successors.

SEPT 12, 2026 · ANTHROPIC

Amodei: "We Must Pace The Frontier"

Anthropic's CEO says AI has advanced drastically faster since summer as models increasingly help build their successors, and commits Anthropic to give outside evaluators permanent, employee-level access.

6-12 mo

window in which an uncontrolled AI botnet could take over the internet, per Amodei

1-2 yrs

timeline Amodei gives for major interpretability breakthroughs

>10%

extinction-risk estimate cited by an Anthropic safety researcher

Time windows named in the essay

6-12 mo

botnet risk

1-2 yrs

interpretability

3-5 yrs

geopolitical window

The recursive self-improvement loop

AI models help build the next generation of AI

→

Capabilities compound faster than before

→

Oversight and safety work struggle to keep pace

A three-part pacing plan

"We must slow the pace at which we improve the capabilities of AI models."

Step 1 (Anthropic, now): permanent, employee-level access for independent evaluators

Step 2: shared safety standards among democratic AI labs

Step 3: global coordination, including authoritarian states

Welcomed

Seen as a concrete, verifiable step, unlike a vague safety pledge

Criticized

A unilateral move; no other frontier lab has matched it so far

Continue reading

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