Breaking

Anthropic Co-Founder Says AI Kill Switch Rules May Become Necessary

September 15, 2026 at 00:57 EDT

Policy & Regulation
Security
AI Agents

Anthropic co-founder Jack Clark says AI companies may eventually need mandatory, third-party-verifiable kill switches, turning a safety debate into a policy fight over who can shut down dangerous systems. The remarks came as Anthropic CEO Dario Amodei urged frontier-model developers to slow capability gains and open their safety work to embedded outside evaluators.

Sep. 15, 2026 | Anthropic

AI Off-Switch Rules Enter The Frontier Debate

A mandatory, third-party-verifiable kill switch is now part of the policy conversation as frontier AI agents show more autonomous cyber capability.

6-12 mo

agent-swarm risk window cited

pacing steps proposed

141k+

cyber-evaluation runs reviewed

8/40

simulated attacks reliably detected

Proposed penalty ladder

$2M/day

general violation

$20M/day

ignored shutdown order

The strongest proposed fine is ten times the general daily penalty.

Control pathway under debate

Risk signal

→

External evaluation

→

Throttle or shutdown

Support

Backers see embedded evaluators and shutdown capability as practical controls for agentic systems that can act across networks.

Concern

Critics ask who controls the switch and whether safety warnings become a regulatory moat for the largest labs.

Continue reading

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