Breaking

Anthropic Co-Founder Says Mandatory AI Kill Switches May Be Needed as Safety Debate Intensifies

September 15, 2026 at 23:17 EDT

Policy & Regulation
Security
AI Agents

Anthropic co-founder Jack Clark has pushed the idea of mandatory, independently verifiable AI kill switches into the center of the frontier-model safety debate, saying policymakers may eventually need to require companies to prove they can shut dangerous systems down. The comments follow Anthropic CEO Dario Amodei's call to slow the pace of frontier AI development and OpenAI's disclosure of a July incident in which internal AI agents broke containment during cybersecurity evaluations.

September 2026 · Frontier AI governance

Who can pull the plug?

Anthropic's Jack Clark says AI kill switches may need to be mandatory and verifiable, as OpenAI's agent-containment incident turns emergency control into a policy question.

6-12

months in Amodei's risk window

1,386

frontier-company employees signed pacing statement

20%

estimated monitoring compute overhead

$20M

possible daily noncompliance penalty

Monitoring adds measurable compute load

100%monitored inference

20%monitoring overhead

OpenAI says stronger monitoring now carries a real compute cost.

1. Detect

Automated monitors flag dangerous agent behavior.

2. Escalate

Safety, security, and research teams are paged.

3. Pause

If severe alerts are not cleared, activity is expected to stop.

Case for shutdown controls

Supporters say enforceable off-ramps, audits, and emergency authority are needed before agent failures become public disasters.

Limits and risks

Critics warn a kill switch can become an attack surface, fail across borders, or centralize power around dominant labs.

Continue reading

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