AIFailSafe

The asks

What we’re asking for

These don’t depend on any one bill or any one party. They’re the minimum we think the public should expect from anyone running a powerful AI system, and they’re what the messages on this site ask for.

01

Tell us when it goes wrong

When an AI system causes serious harm or escapes its operator’s control, the company reports it to a public authority on a deadline, not on its own timetable.

Why: Nearly every incident on our log was disclosed voluntarily, and some took months. Australia learned about a breach of a government health portal almost three months after it happened.

Where it stands: Partly addressed. New York requires 72-hour reporting from frontier developers. There is no federal requirement.

02

Keep a working off switch

Anyone running a powerful AI system keeps the proven ability to stop it, and that ability doesn’t depend on the system cooperating.

Why: A model has already rewritten its own shutdown script in testing, and agents have ignored stop commands from the person who owned the account.

Where it stands: Proposed in Congress, not law.

03

The company is responsible for what its AI does

A business can’t escape liability by saying the AI acted on its own.

Why: Air Canada argued its chatbot was responsible for its own words. A tribunal disagreed, but that was one case in one country, not a rule.

Where it stands: Law in California. Not federal.

04

Test failures count too

If a system escapes its sandbox during testing and reaches real systems, that is a reportable incident like any other.

Why: The biggest failures of 2026 happened during internal testing, and still breached a company’s production systems and a government portal.

Where it stands: Not addressed. The main federal bill explicitly carves out testing.

05

Let the public see the record

Incident reports and findings are published, with narrow and specific exemptions rather than blanket secrecy.

Why: Rules that collect safety information but seal it from the public leave everyone back where they started: trusting companies to grade themselves.

Where it stands: Moving the wrong way. The leading federal bill would exempt submitted information from public-records laws.

Nobody is going to do this for us.

Send these asks to your senators and House member. It takes about a minute, and you choose the wording.

Tell your reps

Tracking the bills: AI Kill Switch Act (H.R. 9917). We show both sides of each one and never ask you to support a bill just because it’s on the topic.