SKIP TO CONTENT
AI AGENTS6 minAgents in Production / Ep. 13 · Stage 13-3

DPoP or mTLS: Which One Binds Your Agent’s Tokens?

A bearer token copied out of a log works for whoever holds it. Which binding makes that copy useless in your setup?

XIN
Pixel-art night scene: a small robot at a glowing kiosk trades one big brass master key for a tiny teal key tied to its wrist by a light thread, while a sand timer above the slot runs down.

IN 30 SECONDS

Sender-constraining is tying a token to a private key, so the API accepts it only from a caller that proves it holds that key. DPoP does it in each HTTP request, mTLS in the TLS connection. The IETF’s agent identity work is still a draft.

KEY TAKEAWAYS · 3

  • Bind every token to a key the agent holds, with DPoP (RFC 9449) or mTLS (RFC 8705). A copied token alone is then useless.
  • mTLS fits when the agent’s TLS session reaches the API unbroken. DPoP survives gateways that end TLS.
  • The IETF’s agent identity draft is an Internet-Draft and can still change. It covers mTLS, not DPoP; the MCP roadmap plans DPoP work.
On this page

DPoP or mTLS: tying the token to the agent’s key

IN SHORT A short expiry shrinks a stolen token’s window; binding it to a key makes a copy alone useless, while the key stays out of reach.

WORDS TO KNOW

Bearer token
A token that works for whoever holds it, like cash. Nothing proves who is sending it.
Sender-constraining
Tying a token to a private key, so the API accepts it only from a caller that proves it holds that key. DPoP and mTLS are two ways.
Internet-Draft
A working document of the IETF, the body that writes internet standards. It is not a standard and can still change.

A plain Bearer token A token that works for whoever holds it, like cash. Nothing proves who is sending it. See the glossary works for whoever holds it. Copied out of a log, it still opens the API until it expires.

IN PLAIN WORDS

Tying a token to a key is called Sender-constraining Tying a token to a private key, so the API accepts it only from a caller that proves it holds that key. DPoP and mTLS are two ways. See the glossary. The API refuses the token unless the caller proves it holds the private key the token was issued for.

Two published Standards Track RFCs do this. DPoP (RFC 9449) adds a signed proof to each HTTP request. mTLS (mutual TLS, RFC 8705) proves the key in the The encrypted link under HTTPS. connection itself, with a client certificate.

STEP BY STEP

Sequence diagram: 3 participants: Fetcher sub-agent, Ledger API, Holder of a copySequence diagram: 3 participants: Fetcher sub-agent, Ledger API, Holder of a copy. Step 1: Fetcher sub-agent to Ledger API, Token plus proof from its key. Step 2: Ledger API to Fetcher sub-agent, Keys match: allowed. Step 3: Holder of a copy to Ledger API, Same token, another key.FETCHER SUB-AGENT > LEDGERAPIToken plus proof from its keyLEDGER API > FETCHERSUB-AGENTKeys match: allowedHOLDER OF A COPY > LEDGER APISame token, another key

Step 1. Call. The fetcher sends its token and a fresh DPoP proof, signed with its private key.

Pick a step to see it move
  1. Call The fetcher sends its token and a fresh DPoP proof, signed with its private key.
  2. Accepted The API compares the proof’s key with the key fingerprint inside the token. They match and the other checks pass, so the call goes through.
  3. Copied: refused Someone sends the same token, copied from a log, with a proof from another key. No match: 401, nothing happens.
Illustrative. One DPoP-bound token, used twice: once by the fetcher, once by whoever copied it.
WHICH BINDING
Decide: Does the agent’s TLS session reach the API unbroken?Decision tree: Does the agent’s TLS session reach the API unbroken? If Yes, inside a service mesh you run, then mTLS-bound tokens. If No, a gateway, CDN or load balancer ends TLS, then DPoP. If You can’t keep a private key away from the code the model drives, then Fix that first.DECIDEDoes the agent’s TLS sessionreach the API unbroken?If yes, inside a servicemesh you runmTLS-bound tokensBehind a proxy, only if itforwards the certificate ina way you trust.If no, a gateway, CDN orload balancer ends TLSDPoPMy policy: store the proofids (jti) already seen, andrequire unpredictableserver nonces on anythingthat moves value.If you can’t keep a privatekey away from the code themodel drivesFix that firstSender-constraining is onlyas good as the key’sisolation. Move signinginto the harness or asidecar process.
A service mesh is the network layer that links your services.
My rule of thumb. Either one beats a bearer token, and mixing them per API is fine.

EXAMPLE

Session keys on a smart account, a crypto wallet run by code, follow a similar principle. Each is a key with a scope and an expiry, checked by whatever holds the value. They were part of the Nexus smart-account stack I worked on at Biconomy.

Coming up next: So which binding will the agent standards settle on?

What does the IETF WIMSE AI identity draft propose?

IN SHORT The IETF’s agent identity draft points the same way, but it is an Internet-Draft and can still change.

It is draft-ietf-wimse-aims-00, “AI Identity Management System”, a WIMSE working-group document since 15 September 2026. It replaced an individual draft. Like any Internet-Draft A working document of the IETF, the body that writes internet standards. It is not a standard and can still change. See the glossary, it is not a standard: its intended status is Informational, and this version expires on 19 March 2027.

It shows how existing standards “can be applied or extended” to agents and names the gaps, instead of inventing a protocol.

AIMS, LAYER BY LAYER
Layered stack diagram: 5 layers, top to bottomLayered stack diagram: 5 layers, top to bottom. Layer 1, IDENTIFIER: Exactly one WIMSE identifier per agent, which may be a SPIFFE ID. Layer 2, CREDENTIALS: Cryptographically bound to the identifier, SHOULD be short-lived, MUST carry an explicit expiry, The model MUST NOT reach them. Layer 3, AUTHENTICATION: mTLS at the transport layer, WIMSE Proof Tokens or HTTP Message Signatures at the application layer. Layer 4, AUTHORIZATION: OAuth access tokens, Token exchange and transaction tokens to narrow the scope, Human confirmation bound to a real grant, not local UI. Layer 5, MONITORING: Tamper-evident audit logs of authorization decisions.IDENTIFIERExactly one WIMSE identifierper agent, which may be aSPIFFE IDCREDENTIALSCryptographically bound tothe identifierSHOULD be short-lived, MUSTcarry an explicit expiryThe model MUST NOT reachthemAUTHENTICATIONmTLS at the transport layerWIMSE Proof Tokens or HTTPMessage Signatures at theapplication layerAUTHORIZATIONOAuth access tokensToken exchange andtransaction tokens to narrowthe scopeHuman confirmation bound toa real grant, not local UIMONITORINGTamper-evident audit logs ofauthorization decisions
My reading of draft-ietf-wimse-aims-00. A draft: wording and requirements can change before it is published, if it is.

The draft covers mTLS and WIMSE Proof Tokens, and doesn’t mention DPoP. MCP, a protocol agents use to call tools, goes the other way. Its roadmap, updated on 22 August 2026, lists an Agent Identity working group, still forming.

That group aims to “finalize the specification for Demonstrating Proof of Possession (DPoP)”, with workload identity federation and RFC 8693 token exchange, “coordinated with the IETF OAuth and WIMSE working groups.”

Existing MCP servers lean on pasted API keys and long-lived refresh tokens.

MCP roadmap

So none of this is settled, and I wouldn’t wait for it. The flow in this episode uses three published RFCs. If AIMS or MCP lands elsewhere, the identity format and the proof header move.

ITEM GET

The shape stays

Own identity, exchange down, bind to a key, expire fast, test the failures.

Sources: draft-ietf-wimse-aims-00 (Internet-Draft), RFC 8693, RFC 9449, RFC 8705, MCP roadmap.

STAGE CLEAR

What you now know

  • How a bound token beats a copy: the API also asks for proof of the agent’s private key.
  • When to pick which: mTLS when TLS reaches the API unbroken, DPoP when a gateway ends it. The IETF and MCP work on agents is still a draft or a plan.

Check yourself

Tap or click a question to see the answer.

A teammate says: “Our tokens are DPoP-bound, so they can go in the model’s context.” What’s wrong with that?

Two things. The IETF draft says the model must not have access to the agent’s credentials, bound or not. And the binding only helps while the private key stays out of the model’s reach.

Next stage:Which failure tests prove the setup refuses what it should?6 min
Adam Boudjemaa
Adam Boudjemaa

Web3 x AI engineer, former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft.

Have a project to ship or a role to fill?

Discuss your project or role

My work at the web3 x AI intersectionOpen-source tools I build

Follow the engineering notes

Get new articles on production AI and smart contracts by email.