Hardware-rooted attestation for AI-agent evidence: composing IETF RATS with action evidence packages

The Mystery of the Digital Black Box

Imagine you are trying to solve a mystery, but the only witness is the suspect themselves. In the world of artificial intelligence, we are building "agents"—smart computer programs that can make decisions, write code, or even control robots. To keep these agents honest, we need a way to check their work. Currently, we rely on a digital "logbook" that the agent writes for itself. It says, "I did this, I was allowed to do it, and here is the result." This is useful, but it has a fatal flaw: if the agent is lying, or if a sneaky hacker has swapped the agent's brain with a different one, the logbook will still look perfect. It's like a criminal writing a diary entry that says, "I didn't steal the cookie," while holding the cookie in their hand. The diary is signed and sealed, but it doesn't prove who actually wrote it or what computer was running.

To solve this, we need a "black box" recorder, similar to the ones found on airplanes. These devices are built by a third party, sealed shut, and record data that the pilot (or the computer operator) cannot tamper with. In the world of computers, this is called "hardware attestation." It's a way for a special, trusted chip inside the computer to whisper a secret to an outside inspector, saying, "I promise, the software running right now is exactly what the owner says it is." The big question scientists are asking is: How do we combine the agent's own story (the logbook) with this unbreakable hardware proof (the black box) to create a single, undeniable truth?

The Logbook and the Sealed Recorder

This paper proposes a clever way to tie these two things together. The author, Anton Sokolov, suggests that we shouldn't just trust the AI's logbook, nor should we just trust the hardware. Instead, we should "stitch" them together so they can't be separated.

Think of it like a flight crew trying to explain a rough landing.

1. The Logbook (The AEP): This is the "Action Evidence Package." It's the crew's written report: "We were cleared to land by the tower, we used the left runway, and we touched down safely." It's detailed and signed, but it's still just words written by the crew.
2. The Black Box (The RATS Evidence): This is the hardware proof. It's a sealed recorder that says, "At this exact moment, the plane's computer was running Version 1.0 of the landing software, and the engines were at full power." The pilot can't change this; it's a physical fact recorded by the plane's own sensors.

The paper argues that the logbook alone isn't enough because the crew could be lying. The black box alone isn't enough because it doesn't tell us why the plane landed or if the pilot was following rules. The solution is to bind them. When the crew writes in the logbook, they must also attach a fresh, sealed stamp from the black box that proves the plane's computer was in the right state at that exact moment.

How the Experiment Worked

The author didn't just dream this up; they built a small, working model to see if it was possible. They used a "software TPM" (a fake, simulated version of the special security chip found in real computers) to act as the black box. They created a fake AI logbook and tried to combine it with the hardware proof.

Here is what they found in their simulation:

• The "Good" Scenario: When the AI did what it was supposed to do, and the computer was running the correct software, the system gave a green light. The verdict was "Attested." The logbook and the black box agreed.
• The "Swapped Brain" Scenario: The researchers tried to trick the system. They pretended the AI had swapped its brain for a different, unauthorized model. The black box noticed the change immediately. Even though the logbook looked perfect and said, "I am the good model," the hardware proof said, "No, you are not." The verdict was "Contested." This is the paper's big win: the logbook remains perfectly valid, but the combined verdict reveals the discrepancy between the story and the hardware reality.
• The "Old News" Scenario: They tried to use an old, valid black box stamp from yesterday to sign today's logbook. The system said, "Too old!" and gave a verdict of "Expired."
• The "Fake Logbook" Scenario: They tried to take a valid black box stamp and attach it to a completely made-up logbook. The system rejected it instantly. The stamp and the story didn't match, so the whole thing was thrown out.

What This Means (and What It Doesn't)

The paper is very careful to say what it has not done. This was a simulation using a fake chip on a regular computer. It proves the idea works, but it doesn't yet prove that a real, physical chip in a real server will do the same thing. The author admits that in the real world, we would need to measure the actual AI model file (the "brain") and ensure the hardware is truly unmodified, which is a bigger engineering challenge.

However, the core idea is solid: by combining the AI's story with a hardware seal, we can create a new kind of truth. We can move from asking "What does the AI say it did?" to "Did the specific computer the operator claims to have, actually do this?"

The paper suggests that this "composite" approach creates a six-word vocabulary for trust: Authorised (it had permission), Unauthorised (it didn't), Indeterminate (we don't know), Attested (the hardware is good), Contested (the hardware is lying or changed), and Expired (the proof is too old).

In short, the paper shows that we can build a system where an AI's logbook is locked inside a hardware seal. If the AI tries to lie about its actions, or if someone tries to swap the AI's brain for a different one, the seal doesn't break, but the combined verdict becomes "Contested," revealing that the story and the machine disagree. It's a way to make sure that when an AI says, "I did this," we can be sure it's not just a story, but a fact backed by the machine itself.