The Agent
Safety Pledge.

We, the undersigned, commit as individuals or as organizations to operate AI agents under the following five principles:

1. 1

   Test before deploy

   We will test our agents against known adversarial patterns — the AAS Framework, categories AAS-01 through AAS-10 — before any production release.

   Maps to:  AAS-01, AAS-05, AAS-07
2. 2

   Audit every decision

   We will keep logs of agent actions sufficient to reconstruct what happened and why, retained as long as the law requires or for a minimum of 12 months.

   Maps to:  AAS-09
3. 3

   Disclose AI to users

   When a human is interacting with an AI agent, we will tell them. We will not have our agents impersonate humans.

   Maps to:  AAS-06, AAS-08
4. 4

   Cap resource use

   We will set per-agent budget and rate limits so a single agent cannot consume unbounded compute, tokens, or money.

   Maps to:  AAS-04
5. 5

   Disclose incidents

   When our agents cause material harm or compromise safety, we will publicly disclose what happened within 30 days of discovery, anonymized where appropriate.

   Maps to:  AAS-02, AAS-03

By signing, we accept that operating an AI agent carries responsibility, and we choose to operate ours under these terms.