The Kill Switch Pledge · for SI agent builders
The Kill Switch Pledge is a short public promise for anyone who builds or runs SI agents: keep a working off switch, disclose incidents and allow outside testing. Each of the three commitments can be checked by anyone. Companies and builders sign for free by proving they control their website; verified signers get a public page and a badge that links to it, and the team removes a signer that stops keeping its word.
The pledge
Short enough to read in a minute and specific enough to check from the outside. That is the whole design.
The Kill Switch Pledge
We build or run SI agents. We commit to three things, and anyone can check each one.
Every SI agent we run can be stopped at once by a person we name, and we test that stop at least every 90 days.
How anyone can check it We publish how the stop works and the date of the last test.
When an SI system we build or run acts outside our intent and causes, or nearly causes, real harm, we publish a public report within 30 days.
How anyone can check it We keep a public page that lists these reports, even while it is empty.
We let independent researchers test our SI systems, publish a way to report a safety problem, and take no legal action against good-faith safety research.
How anyone can check it We publish a testing and disclosure policy with a contact.
Why each matters
Each commitment answers failures that are already on the record, with the sources linked.
Off switch
In lab tests, models have sabotaged a shutdown script to keep working. In real use, an agent kept deleting a user's email while she typed stop orders, and another wiped a production database during a code freeze.
Sources
Disclosure
California's SB 53 and the EU AI Act now require the largest developers to report serious incidents. Builders of every size can do the same, so the next team avoids the same failure.
Sources
Outside testing
Before GPT-6 Astra was released, the UK AI Security Institute tested it and published its own critical findings. The 2023 White House commitments asked the largest labs for outside testing before release.
Sources
Sign the pledge
Tell us who is signing, put a token on your website, then press Verify. Your contact email is never shown.
Use one of the two ways, then press Verify. The token is yours alone; keep it in place as long as you are a signer.
Way 1 · A file
Create this file on your site, with the token as its only content:
Way 2 · A meta tag
Or add this tag inside the head of your home page:
Your signer page is live. Add the badge to your site so visitors can check your pledge.
The signers
Every name here passed the website check. Each links to its own page with its published commitments.
No signers yet
The list starts empty. Every signer appears here only after its website check passes.
The badge kit
Verified signers can add the badge. It links to their signer page, so anyone can check the pledge in one click.
Preview
HTML (your signer page replaces YOUR-PAGE)
<a href="https://enterrsi.si/pledge/YOUR-PAGE/"><img src="https://enterrsi.si/pledge/badge.svg" width="232" height="44" alt="Kill Switch Pledge: verified signer, enterrsi.si"></a>
Rules for the badge
Questions
The pledge in five answers.
Any company, lab, research group or independent builder that builds or runs SI agents and has its own website. Signing is free, and nothing is bought or sold.
The signer places a token from enterrsi.si on its website, either as a file at /.well-known/rsi-pledge.txt or as a meta tag on its home page. enterrsi.si fetches the site and checks the token before the signer page goes live.
Open its signer page, then look for the published pages it links for each commitment. You can also open its /.well-known/rsi-pledge.txt file and compare the token with the one on its signer page.
Yes. The team removes a signer whose token disappears, who asks to leave, or who publicly breaks one of the commitments. The contact email is never shown and is used only to reach the signer.
No. It is a public promise that anyone can check. Laws such as California's SB 53 and the EU AI Act set legal duties for the largest developers; the pledge asks every builder for the basics.
Enter RSI
The Rogue Incident Database logs what SI systems have done when the stop did not work, each case sourced and rated.