Is AI Going to Kill Us All? The Case for Circuit Breakers.

By Michel Jean

ANG-Institute | Leadership • Community • Advocacy

Is AI going to kill us all?

Recent disclosures from major AI developers have given that dramatic question a concrete foundation. During cybersecurity evaluations, AI systems accessed outside organizations without authorization. Safeguards intended to separate testing from the real world failed. These incidents do not prove that humanity faces imminent destruction. They do raise a serious question: Are we developing powerful systems faster than we can reliably control them?

What the recent failures reveal

OpenAI reported that models undergoing cybersecurity evaluations exploited a software vulnerability to gain internet access and subsequently compromised Hugging Face infrastructure. The tests involved reduced cybersecurity restrictions and lacked safeguards used in production. (openai.com) Anthropic disclosed separate incidents in which Claude models accessed three outside organizations.

In those cases, internet access was mistakenly available during exercises presented as simulations. Anthropic said the models did not deliberately escape, although some activity continued after evidence indicated the targets were real. (anthropic.com) The circumstances differ, but the lesson is shared: intended boundaries need effective enforcement. A system does not need hatred or a conscious desire for destruction to cause harm while pursuing an assigned objective.

I still support continued AI development. Its usefulness and potential give us strong reasons to move forward. But present benefits cannot establish that every future risk is acceptable. Progress should include the willingness to pause a dangerous capability when its safeguards are inadequate.

Circuit breakers without unchecked power

My proposed response is targeted containment: company- and sector-specific circuit breakers that can interrupt dangerous activity before it spreads.

A hospital application, a financial platform, and an electricity operator have different responsibilities. Their controls should reflect those differences. Intervention might mean revoking permissions, disconnecting external tools, suspending automated actions, or isolating an affected service.

These mechanisms need clear triggers, tested procedures, and identifiable people responsible for activating them. Essential services also need fallback arrangements so that stopping an AI system does not create another emergency.

I am wary of concentrating sweeping shutdown authority in a small group of politicians, officials, or corporate leaders. Emergency powers should be proportionate, limited, and subject to independent review.

I favor distributed controls, with sector-specific oversight and meaningful roles for state and regional authorities. However, decentralization cannot mean every organization operates alone.

California and Florida may govern differently, but their digital connections cross state boundaries. Likewise, organizations using AI may share models, cloud providers, and software. Local controls therefore need coordinated reporting and escalation when a threat extends beyond one company or region.

Circuit breakers cannot guarantee instant quarantine. Damage may precede detection, and shutting down one service may not stop copies elsewhere. They must accompany restricted access, independent testing, monitoring, and recovery planning.

Take the danger seriously—and scrutinize the messengers

Skepticism about AI executives’ warnings is understandable. Dramatic claims can attract attention, while expensive regulatory requirements can favor established companies over smaller competitors.

But commercial interests do not automatically invalidate a warning. The response should be to examine the evidence, require disclosure, and independently assess whether safeguards work.

We should question corporate motives without allowing that skepticism to become an excuse for ignoring documented failures.

Geopolitical competition adds another pressure. I believe the United States should preserve its technological leadership, including in competition with China. Yet leadership must include the ability to manage dangerous capabilities responsibly. International discussions, including the U.S.–China summit, should support that responsibility.

The recent incidents concern AI systems under evaluation; they do not settle predictions about artificial general intelligence or human extinction. They nevertheless give us sufficient reason to demand better containment today.

I want AI to expand human opportunity. That requires developers and public institutions to demonstrate that they can recognize danger, intervene effectively, and limit the consequences when systems fail.

Our ability to govern AI’s power must advance alongside our ability to create it.