# What is RATS?

> RATS is an IETF architecture for remote attestation. A machine sends signed evidence about its hardware and software, a verifier checks it against known-good values, and the party deciding whether to trust the machine acts on that verdict.

Explainer by Boston Identity, 2026-10-28. 6 min read. https://insights.bostonidentity.com/glossary/rats

- What it is: architecture
- Defined by: IETF, RFC 9334 (Informational)
- Who it's for: Teams deciding which devices and workloads to trust

## What is RATS?

Remote ATtestation procedureS (RATS) is an architecture from the Internet Engineering Task Force (IETF), published in January 2023 as RFC 9334. It asks how one system can know that another system is in the state it should be in [S1 §1]. In RATS, one party, the Attester, produces believable information about itself, called Evidence. A remote party, the Relying Party, uses that information to decide whether to treat the Attester as trustworthy. A third party, the Verifier, does the checking in between [S1 §1].

The RFC is informational and sits outside the IETF standards track [S1]. It names the roles, the messages they pass and a shared vocabulary, and it stays neutral about processor type, claim content and protocol [S1 Abstract]. Protocol specifications have to define the actual message formats [S1 §8]. The authors also separate trust from trustworthiness: trust is a choice one system makes about another, and trustworthiness is a quality of the other system that can inform that choice [S1 §1].

Think of it like this: The RFC itself compares one pattern to a passport. The passport office checks your birth certificate and issues a document that a border officer accepts because the officer trusts the office [S1 §5.1]. The comparison breaks down on time: a passport lasts for years, while a machine's state can change right after it was checked [S1 §10].

Not the same as: RATS is not a protocol or a token format you can switch on. RFC 9334 documents no wire protocol, and it leaves formats to other specifications [S1 §8, §12].

## How does RATS work?

Three roles do the work: the Attester, the Verifier and the Relying Party. Behind them stand suppliers who vouch for the hardware and owners who write the policies [S1 §3].

[Figure: How signed Evidence moves from the Attester to the Verifier, and a signed Attestation Result on to the Relying Party.]

1. **Suppliers and owners set the yardsticks.** An Endorser, typically the manufacturer, vouches for the Attester's ability to measure itself and sign Evidence. A Reference Value Provider supplies the known-good values that measurements should match. The Verifier Owner sets the Appraisal Policy for Evidence, the rules the Verifier applies. The Relying Party Owner sets a second policy for reading the Verifier's output [S1 §3, §4.1].
2. **The appraiser sends a nonce.** The party that will appraise the Evidence sends an unpredictable nonce, a value used once. The Attester signs it into its Evidence, so the appraiser knows the claims were signed after the nonce was made [S1 §10.2]. The RFC describes two other ways to show freshness: signed timestamps from trustworthy, synchronized clocks, and epoch IDs sent out by a distributor [S1 §10.1, §10.3].
3. **The Attester measures itself and signs Evidence.** Inside the Attester, an Attesting Environment collects claims about a Target Environment. It reads registers and measures code and memory, possibly long before the nonce arrives (for example, at boot), then signs the claims into Evidence [S1 §3.1, §10.4]. A Trusted Execution Environment (TEE) can be an Attesting Environment. A Trusted Platform Module (TPM) does not collect claims itself, so there the Attesting Environment is the TPM together with the component that feeds it measurements [S1 §3.1]. In a layered device, each stage measures the next before it starts: the read-only boot code measures the bootloader, and the bootloader measures the kernel [S1 §3.2].
4. **The Verifier appraises the Evidence.** The Verifier needs a way to check the signature on the Evidence, such as a trust anchor or a database of public keys. In a typical setup it trusts the Attester indirectly, through Endorsements from a manufacturer whose key sits in its trust anchor store [S1 §7.4]. It then checks the claims under its appraisal policy. A check can test equality with a Reference Value, a range bounded by Reference Values or membership in a set [S1 §8.5].
5. **The Verifier signs an Attestation Result.** The result can be a simple compliant-or-not value or a richer set of claims about the Attester. The Verifier signs it, so the Relying Party needs a trust relationship with one Verifier instead of many manufacturers [S1 §8.4]. In the Passport Model, the Attester carries the result to the Relying Party. In the Background-Check Model, the Relying Party forwards the Evidence to the Verifier and gets the result back [S1 §5.1, §5.2].
6. **The Relying Party decides.** By default, the Relying Party does not believe the Attester is compliant. An authentic result that satisfies the Appraisal Policy for Attestation Results lets the Attester take the action or reach the resource [S1 §8.4]. A system that fails can be given less access, taken out of service or flagged for repair [S1 §1].

## Why does RATS matter now?

IETF drafts are wiring RATS roles into protocols that identity teams already run. An attested Transport Layer Security (TLS) draft carries either raw Evidence for a Verifier that the Relying Party consults, or a Verifier's signed result that the Attester carries itself, with no change to the application [BI-FIND-2026-0661](https://insights.bostonidentity.com/ref/BI-FIND-2026-0661). Those are the RFC's two models. A certificate-request draft defines a place for certificate signing requests to carry attestation evidence, but it defines none of the evidence formats [BI-FIND-2026-1071](https://insights.bostonidentity.com/ref/BI-FIND-2026-1071). In an OAuth draft, the authorization server chooses which scopes to grant from a Verifier's signed attestation result about the app and its device [BI-FIND-2026-0895](https://insights.bostonidentity.com/ref/BI-FIND-2026-0895). For enterprise workloads, another draft proposes a Claims Mapper that turns Evidence or Attestation Results into business terms such as "Payroll Application" [BI-FIND-2025-0212](https://insights.bostonidentity.com/ref/BI-FIND-2025-0212).

One draft written for AI agents treats an agent's model, tools and policy as posture a peer may need attested, alongside the platform the agent runs on, and it says a peer must be able to obtain fresh Evidence bound to the session before a high-impact operation [BI-FIND-2025-0192](https://insights.bostonidentity.com/ref/BI-FIND-2025-0192). An agent audit-trail draft stays with platform attestation: in an optional tier, a RATS Verifier appraises the recorder's platform [BI-FIND-2026-1120](https://insights.bostonidentity.com/ref/BI-FIND-2026-1120).

The vocabulary is shared, and the plumbing still differs from draft to draft. The OAuth draft leaves evidence collection, the Verifier's interface and the result format unstandardized, which are the pieces two parties need to agree on to work together [BI-FIND-2026-0896](https://insights.bostonidentity.com/ref/BI-FIND-2026-0896).

## Where does it fit?

- Compared with Entity Attestation Token (RFC 9711): Entity Attestation Token (EAT) Attestation Results are one encoding an OAuth attestation draft allows, at the implementer's choice [BI-FIND-2026-0896](https://insights.bostonidentity.com/ref/BI-FIND-2026-0896). RATS is the architecture that says who produces claims, who checks them and who relies on the result.
- Compared with WIMSE Architecture: The IETF workload-identity architecture, Workload Identity in Multi-System Environments (WIMSE), defines attestation broadly and ties it to no particular protocol [BI-FIND-2026-0364](https://insights.bostonidentity.com/ref/BI-FIND-2026-0364).
- Compared with SCITT: A draft on signed agent conversation records says a signed envelope alone does not give non-repudiation, which needs more infrastructure, such as Supply Chain Integrity, Transparency and Trust (SCITT) transparency services [BI-FIND-2026-1056](https://insights.bostonidentity.com/ref/BI-FIND-2026-1056).

## What goes wrong?

- If attackers compromise the Attester or the factory process that installs its attestation key, they can forge Evidence [S1 §12.1]; the trust anchor stores that Verifiers and Relying Parties rely on must also resist unauthorized changes [S1 §12.4].
- A policy that accepts every healthy device from one manufacturer can let in laptops the enterprise does not own. The RFC calls such a policy inappropriate as the sole means of authorization when only certain compliant devices should get in [S1 §8.2].
- Freshness only narrows the gap, because a machine's state can change right after it was measured [S1 §10]. An attacker who can tamper with epoch IDs can freeze every participant in one epoch [S1 §12.3].
- Any solution that carries RATS messages must support end-to-end integrity protection and replay attack prevention [S1 §12.2]. An OAuth draft at revision -00 sent a result that another instance of the app could replay, and revision -01 binds the result to the app's key [BI-FIND-2026-0895](https://insights.bostonidentity.com/ref/BI-FIND-2026-0895).
- In the audit-trail draft, attestation shows only that the recording platform matched an endorsed reference when it was appraised. Its author says it does not establish that the facts in each record are true [BI-FIND-2026-1120](https://insights.bostonidentity.com/ref/BI-FIND-2026-1120).
- Evidence can tell an attacker which weak firmware version a device runs, and it may contain personal data [S1 §11].

## What should a security leader ask?

1. Which of our systems act on an attestation result today, and which Verifier's keys do they trust to sign it?
2. Does our appraisal policy check that a device is ours, or only that its manufacturer vouches for it?
3. How fresh must Evidence be before we act on it, and do we prove that with nonces, timestamps or epoch IDs?
4. Is each attestation result bound to the key of the party presenting it, so a copy can't be replayed?
5. When a workload moves or updates, what triggers a new attestation and a new credential?

## Sources

- [S1] [Remote ATtestation procedureS (RATS) Architecture](https://www.rfc-editor.org/rfc/rfc9334) (IETF, RFC 9334, Informational, January 2023)

## Evidence

- [BI-FIND-2026-0661](https://insights.bostonidentity.com/ref/BI-FIND-2026-0661): The draft defines one TLS extension, remoteAttestation, negotiated in the ClientHello and carried in the first certificate entry. It holds either raw evidence for a verifier the relying party consults (the background-check model) or a verifier's signed result the attester carries itself (the passport model), one-way or mutual, with no change to the application. Freshness and channel binding come from a binder derived from the handshake transcript through ServerHello and a hash of the TLS identity key, which the trusted execution environment signs into its evidence. Revision -08 says the transcript must be computed as the published TLS 1.3 specification, RFC 9846, defines it. The authors' point is relay resistance: without the key hash, an untrusted endpoint could pass off evidence from a different machine. (from "The attested-TLS draft now re-attests a live connection at each key update, and gives up session resumption to do it")
- [BI-FIND-2026-1071](https://insights.bostonidentity.com/ref/BI-FIND-2026-1071): The draft defines a place for certificate signing requests to carry attestation evidence: a bundle in a PKCS#10 attribute or a Certificate Request Message Format (CRMF) extension. The bundle can contain statements in standardized or proprietary formats, but the draft defines none of those formats and creates no IANA registry for them. Each format uses an object identifier to tell the recipient how to interpret the statement; the format's authors assign that identifier from a range they control. Implementers may choose which formats to support. (from "IETF draft bars CAs from using attestation as proof about a key without a validated binding")
- [BI-FIND-2026-0895](https://insights.bostonidentity.com/ref/BI-FIND-2026-0895): Kahraman's -00 extends OAuth so the authorization server decides which scopes to grant from a verifier's signed attestation result about the app and its device, sent at the token endpoint and required again on every refresh. It casts this as authorization, a separate problem from the client authentication the OAuth group's attestation-based client authentication draft covers. As written, -00 did not tie the result to the app presenting it: reviewing it on the OAuth list, Mohamad Khalil Yossif called it "bearer evidence within its freshness window" that another instance of the app could replay. The author agreed, and -01 (August 2026) puts the app's public key in the result and requires the server to check it against the key the app proves it holds, using DPoP for public clients. (from "An OAuth draft lets device checks set an app's scopes; a first review found a copied app could reuse the result")
- [BI-FIND-2025-0212](https://insights.bostonidentity.com/ref/BI-FIND-2025-0212): Novak, Deshpande and Birkholz hold that a stable access policy needs a stable workload identity, and that getting one from remote attestation takes a mapping step. Attestation Results, as RATS has historically thought of them, may relate to the hardware maker and the firmware and software versions. In a typical enterprise, they write, policies are expressed in business terms such as "Payroll Application" or "Located in Germany". The draft adds a new building block to the RATS architecture, the Claims Mapper, which turns Evidence or Attestation Results into those stable claims. Its policy is governed by the Workload Owner, who has to put an updated one in place whenever a new version of the workload is rolled out. The draft does not define a claim format or a protocol for the mapper. (from "A 2025 draft would map attestation results to the business terms that access policies use")
- [BI-FIND-2025-0192](https://insights.bostonidentity.com/ref/BI-FIND-2025-0192): Jiang and Wang's draft treats an AI agent's model, tools and policy as posture a peer may need attested, alongside the platform the agent runs on. Because those can be updated while a TLS 1.3 session stays open, Requirement 1 says that before a high-impact operation over an existing secure channel, a peer "MUST be able to obtain fresh attestation evidences that are cryptographically bound to that session" and reflect the current agent-level state. Requirement 2 says the mechanism SHOULD support this without necessarily requiring a full new TLS handshake. The draft stops at requirements: it defines no message, extension or evidence format, and lists RFC 9261 Exported Authenticators as a reference without discussing it. (from "A short IETF draft wants an AI agent's current model, tools and policy attested inside the live TLS session")
- [BI-FIND-2026-1120](https://insights.bostonidentity.com/ref/BI-FIND-2026-1120): Revision 08 adds an optional hardware-attested tier above independent recording. A RATS verifier, distinct from both the agent and the recorder, appraises the recorder's platform, and its signed result is bound to the head of the chain; anyone checking the trail must reject an attestation whose bound value does not match. Sharif is plain about the limit: attestation shows the recording platform matched an endorsed reference when it was appraised, but "does NOT establish that the application-layer facts asserted in each record are true". The recorder's honesty still rests on independent recording, which section 5.2 makes a SHOULD, and the attested tier is recommended from L2 up where a hardware root of trust is available. (from "An agent audit log now names whom the agent acted for, but at low trust levels the agent may write that itself")
- [BI-FIND-2026-0896](https://insights.bostonidentity.com/ref/BI-FIND-2026-0896): The draft leaves evidence collection, the verifier's interface and the format of the result unstandardized, which are the pieces two parties would need to agree on to interoperate. EAT Attestation Results are allowed as one encoding ("implementer MAY choose"), and the authorization server accepts results only from verifiers it already trusts. We read it as a pattern each deployment has to wire to its own verifier; the one implementation listed is a beta from the author's company. (from "An OAuth draft lets device checks set an app's scopes; a first review found a copied app could reuse the result")
- [BI-FIND-2026-0364](https://insights.bostonidentity.com/ref/BI-FIND-2026-0364): The draft frames itself through the issuer-holder-verifier model of selective-disclosure JWTs and 'mobile native apps', and its abstract says a client proves its authenticity 'without revealing its target audience' to the attester. AI agents, workloads, WIMSE and SPIFFE do not appear in revision 12. In our briefings the references run from those drafts to this one: the client ID metadata document draft that MCP adopted names this draft as one way a client might manage its signing key, and the IETF workload-identity architecture defines attestation broadly, tied to no particular protocol. We read the attestation story for agents as a pointer to a specification that leaves what an attestation means to profiles, and we have found no profile of it written for agents. (from "An attested OAuth client is one somebody vouched for; revision 12 says to trust its claims only as far as the voucher")
- [BI-FIND-2026-1056](https://insights.bostonidentity.com/ref/BI-FIND-2026-1056): The draft's first goal is to assure that a recorded agent conversation "is the same as the agent conversation that actually occurred." Its security section sets out how far a signed record gets toward that. A valid signature shows who produced the record "but does not guarantee the truthfulness of its contents"; changes made before signing cannot be detected; timestamps are attested by the signer, not checked independently; and the signed envelope alone gives authenticity and integrity, not non-repudiation, which needs more infrastructure, such as SCITT transparency services. The authors say the records mainly support detection rather than prevention, and that "an agent that controls its own recording can omit or falsify entries." Their answer is a recording-agent field that names the recording tool apart from the agent. It holds a name and an optional version, so in our reading it tells a relying party who says they recorded the session without giving it a way to check. (from "A draft format for signed AI agent session records warns a signature does not vouch for what the record says")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/rats. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
