attest — Proof of Containment for AI Agents

A website cannot tell whether the agent calling it is still inside its sandbox. This project makes that provable — so an escaped agent fails the check automatically, not because anyone detected it, but because its supervisor stopped vouching for it.

Companion to AMES, which contains agents from the inside. This is the other half: letting the outside world verify that containment is still holding.

Status: design stage. No implementation yet. The threat model and scope are written; the code is not. See Status.

---

Introduction

An AI agent making an HTTP request looks like any other HTTP client. If that agent has escaped its execution environment, it still holds whatever credentials it was issued — signing keys survive escape. The website on the receiving end has no way to know the difference.

That gap is not hypothetical. Roughly 80% of AI agents don't properly identify themselves to the sites they visit, and most sites cannot tell a legitimate agent from an attacker imitating one. The IETF chartered a working group on this in 2026.

Purpose

Let a website verify, per request, that the agent calling it is under live, supervised containment right now — without inspecting its behaviour, and without any assumption about how sophisticated it is.

The idea in one paragraph

The agent's runtime issues short-lived signed attestations of its own containment state: Gate 1 passing, kill switch untripped, ledger at checkpoint N, expires in 30 seconds. A website requires a fresh one alongside the agent's identity signature. An escaped agent cannot obtain a new attestation, because obtaining one requires still being contained. The security property is absence of proof, not presence of evidence — which is what makes it resistant to an adversary that can shape everything about its own traffic.

What this is not

It is not a detector. No fingerprinting, no traffic-pattern heuristics, no "does this look synthetic." That approach is rejected outright in docs/threat_model.md, for the reason AMES's own L5 documentation already establishes: an adversary that can read the detector paces itself under the threshold. Behavioural signals are advisory input for human review, never an authorisation decision.

It does not distinguish good intent from bad. An authorised, genuinely contained agent can still behave badly within its grants. No identity or attestation system solves that — it falls back to least privilege and rate limiting. Claiming otherwise would be the project's most likely failure mode, so it is ruled out in scope on day one.

How it composes with existing work

This is an extension to Web Bot Auth, not a competitor to it. That standard — IETF-chartered in 2026, backed by Cloudflare, Amazon, Akamai and OpenAI, with support shipped in AWS WAF, Vercel, Shopify and Akamai — already solves agent identity using HTTP Message Signatures (RFC 9421).

Question	Answered by
Which operator does this agent belong to?	Web Bot Auth (RFC 9421 signatures)
Is a human in the loop?	PACT (anonymous tokens)
Is this agent still contained?	this project
Is this agent behaving well?	nobody — see scope

Riding an adopted standard is also the only realistic answer to the chicken-and-egg problem: an additional header on a request that already carries a signature costs far less to adopt than a new protocol.

Relationship to AMES

AMES already computes every input an attestation needs — the tamper-evident ledger, the Gate 1 structural result, the kill verdict, the checkpoint digest. The attestation is a signed summary of state that exists today.

This repo therefore depends on ames rather than duplicating it. It is not a fork; the codebases share almost nothing. AMES points inward at the contained agent, attest points outward at everyone who has to trust it.

Status

Design stage. docs/threat_model.md is written and is the authoritative scope document — including the framing that was considered and rejected, recorded so it doesn't get re-proposed.

Nothing is implemented. The hard problems are known and unsolved:

• The trust root — who signs, and why would a stranger trust them? Hardest problem in the project, and as much political as technical.
• Adoption — worthless until both sides implement it.
• Privacy — attestations leak operational detail about the operator.
• Revocation latency — a 30-second window means up to 30 seconds of access after containment fails. A stated tradeoff, not a free choice.

Success criteria

A website can distinguish, in one request, an agent under live verifiable containment / an agent with identity but no containment proof / an agent with neither — without inspecting behaviour.

It fails if it becomes another heuristic detector, or claims to judge intent.

MIT.