SCITT
Supply Chain Integrity, Transparency and Trust · SCITT working group
By Boston Identity · Explained · 6 min read
The short answer
SCITT is an IETF architecture for a tamper-evident record of signed claims about software. A transparency service checks each signed statement, adds it to an append-only log and returns a receipt that proves the statement is there.
- What it is
- An architecture
- Defined by
- IETF SCITT WG; draft -22, later RFC 9943
- Who it's for
- Teams that sign, ship or audit software and agent records
- In our briefings
- 301by month, last 12
What is SCITT?
SCITT stands for Supply Chain Integrity, Transparency and Trust. It is the name of an IETF (Internet Engineering Task Force) working group and of the architecture that group wrote. We read draft -22 of that architecture, a Standards Track draft from October 2025; the IETF later published it as RFC 9943.S1
The problem it solves is accountability for claims about software. A vendor, a build system or an outside reviewer makes a statement about an artifact: a software bill of materials (SBOM), a scan result, a release approval. The draft wants the makers of such statements held to account, so they cannot "hide or shred the evidence when it becomes inconvenient later".§2.1 The architecture checks that a statement is signed and that its author can be identified, and it gives everyone the same history of that statement. It leaves the syntax and meaning of each kind of statement to the people who use them.§2.1
The draft calls itself a generalization of Certificate Transparency, which the draft reads as a transparency architecture for the supply chain of X.509 certificates.§4 SCITT applies the same idea to any signed statement, and the draft uses the software supply chain as its first example.§1
A SCITT signed statement on its own is not proof of registration. The signature shows who made the statement; only a receipt from a Transparency Service shows the statement was registered there.§4
How does SCITT work?
An Issuer signs statements, a Transparency Service keeps the log, and a Relying Party checks the result before it trusts an artifact. A Client, which need not be the Issuer, carries statements to the service.
Figure steps
- The Issuer signs a statement
- A Client submits it for registration
- The service checks the signature and its policy
- The service appends it to the log
- The service returns a receipt
- The Client attaches the receipt
- The Relying Party verifies before it trusts
How a signed statement travels from Issuer to log to Relying Party in SCITT.
Drawn from the SCITT architecture draft, sections 5 to 7
Swipe for the full figure →
The Issuer signs a statement
The Issuer wraps a statement about an artifact in a COSE signature. The protected header must name the Issuer and the subject, the artifact the statement is about.§6 A large or sensitive payload can be replaced by its hash, so the full content need not be sent to the service.§6.2
A Client submits it for registration
A Client authenticates to the Transparency Service and submits the signed statement. The Client may act for one or more Issuers. How it authenticates, and what it is allowed to do, is left to each implementation.§6.3
The service checks the signature and its policy
The Transparency Service must verify the Issuer's signature against a trust anchor it holds, such as a root certificate. It then applies its registration policy, the rules a statement must meet to be accepted.S1 The policy and trust anchors must themselves be registered in the log, so auditors can see which rules applied at the time.§5.1.1.1
The service appends it to the log
The service adds the statement to its verifiable data structure, a log built so that its history can be checked by others. That structure must be append-only, so entries cannot be changed, deleted or reordered. It must also show everyone the same history, which the draft calls non-equivocation.§5.1.3
The Client attaches the receipt
The Client adds the receipt to the signed statement, which makes it a Transparent Statement. The same statement may be registered with several services, and all their receipts can travel together.§6.3
The Relying Party verifies before it trusts
A Relying Party checks the receipt before it trusts the artifact. It must trust the key of at least one receipt issuer. It may verify a single receipt it accepts and skip the statement's own signature.§7.1 Auditors go further: they replay the log to check that the service kept it consistent.§5
Why does SCITT matter now?
SCITT was written for software supply chains, but the drafts we read this year reach for it to log what AI agents do. In one profile for recording agent actions, the operator signs a record of each consequential action, and an independent custodian running a SCITT transparency service holds the chain append-only.1 Another draft logs an agent's authorization decision on a SCITT log before the action runs.2 A third lets a session hash be anchored with a SCITT transparency service.3
The agent identity drafts use it too. In one agent naming draft, the highest verification tier adds a lookup in a public transparency log that must operate as a SCITT Transparency Service.4 The authors of a draft on signed agent conversation records say a signature alone does not give non-repudiation, and they name SCITT transparency services as the kind of infrastructure that would.5
Citing SCITT and conforming to it are separate steps. The agent-action profile's revision -01 does not yet specify its receipts as COSE Receipts, as RFC 9943 requires, and defines no inclusion proof.1 A security leader who hears "logged on SCITT" should ask which parts of the architecture the product implements.
Where does it fit?
| Compared with | How SCITT differs |
|---|---|
| RFC 9943 SCITT architecture | RFC 9943 is the published architecture document. SCITT is the wider effort: the working group, its drafts, and the profiles other authors build on the architecture. |
| COSE | COSE is the format SCITT uses for signed statements and receipts. COSE signs one object; SCITT adds the shared log that proves a signed object was registered. |
| JSON Canonicalization Scheme | JSON Canonicalization Scheme fixes the bytes of a JSON document so two parties hash it the same way. SCITT says nothing about JSON layout; it logs signed statements and issues receipts for them. |
What goes wrong?
- Registering a statement only proves that the Issuer produced it. Issuers can make false statements, on purpose or by mistake.§9.2
- Beyond the mandatory checks, a service's registration policy may add no further checks, even one that allows all authenticated submissions.§5.1.1 A receipt alone says little about how a statement was vetted.
- An Issuer can refuse to register a statement, or submit some and not others. The draft says it is important for relying parties not to accept a signed statement unless they can find a receipt from a service they trust.§9.3
- Unless the service's policy says so, a relying party cannot assume the log's order matches the order in which statements were signed.§9.1
- A service whose receipt signing key is compromised can roll its log back to a point before the compromise and issue fresh receipts with new credentials.§9.4.2
- Independence is up to the deployer. In the agent-action profile, if one entity runs both the operator and custodian roles, the tamper-evidence guarantees are void.1 The authors of the conversation-records draft add that an agent that controls its own recording can omit or falsify entries.5
What should a security leader ask?
- 01
Who runs the transparency service our records go to, and is it independent of the team that writes the records?
- 02
Do our relying parties refuse a signed statement that comes without a receipt from a service we trust?
- 03
Does the product that says it logs on SCITT issue COSE receipts with inclusion proofs, or only a hash chain of its own?
- 04
Whose public key do we trust to check receipts, and how would we learn that it was compromised?
- 05
Which of our statements, such as SBOMs, scan results or agent action records, carry personal or confidential data a service should not hold?
Sources
- S1
An Architecture for Trustworthy and Transparent Digital Supply Chains (draft-ietf-scitt-architecture-22)IETF SCITT Working Group · Internet-Draft -22, October 2025; later published as RFC 9943
›Evidence: 5 claims from 5 briefings
- 1BI-FIND-2026-0972 · A transparency-log profile for agents exposes edited and dropped records; an action nobody logged leaves no gapIn Emirdag and Feichtinger's profile the agent's operator signs a record of each consequential action, and an independent Evidence Custodian running a SCITT transparency service holds the chain append-only. Once a record is admitted nobody can alter or reorder it unnoticed, and a gap in the sequence numbers shows a record produced and not registered. "An action that was never recorded leaves no gap", however. The threat model excludes an operator that records an action inaccurately or not at all, and compromise of the agent. Only the admission time on the custodian's receipt comes from outside the operator; the timestamps inside the record are the operator's own. If one entity runs both roles, "the tamper-evidence guarantees are void", and a colluding operator and custodian can alter evidence without detection by relying parties that trust only that custodian's receipts. Revision -01 has not yet specified its receipts as COSE Receipts, as RFC 9943 requires, defines no inclusion proof or non-equivocation proofs, and withdrew -00's claim that the hash chain matches a Merkle inclusion proof.
- 2BI-FIND-2026-0865 · Munoz's agent Permit is signed before the action runs, but the profile does not make the action wait for itMunoz's Permit records an agent's authorization decision before the action, signed and logged on a SCITT transparency log, but the profile does not make the action wait for it. The draft calls the Permit "an evidence artifact" and says it does not stop an agent "reaching a provider by a path the Issuer never sees"; no step requires the tool or service to check a Permit before running. The verifier's checks run after the fact, including a comparison of the Permit's hash with the digest in a closing record of what was sent, and in the reference code that digest is copied from the Permit, not measured from the request. Only the author's own proxy refuses a denied or altered request before sending it.
- 3BI-FIND-2026-0730 · An agent audit log now names whom the agent acted for, but at low trust levels the agent may write that itselfRevision 07 adds a section setting the draft beside Gilda's WIMSE agent audit record, and Sharif states the split himself: Gilda's work requires attestation by an observer the agent does not control, while AAT permits self-recording at L0 and L1. That section describes AAT as requiring an independent recorder from L2 up, but the normative text in section 5.2 says SHOULD. It treats Munoz's WIMSE authorization-evidence records as one way to fill the principal's grant reference, lets a session hash or batch root be anchored with a SCITT transparency service, and keeps its own schema rather than OpenTelemetry or CloudEvents, which it says lack its chaining, pre-execution semantics and principal binding, though AAT records may be carried inside them.
- 4BI-FIND-2026-0303 · The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted itThe draft's bronze, silver and gold are not grades awarded to an agent. 'A tier describes what the client verified, not a property the RA assigned': bronze means the caller checked the certificate, silver adds a DNS-based check that needs DNSSEC, and gold adds a lookup in a public transparency log that 'MUST operate as a SCITT Transparency Service'. If the log cannot be reached the caller drops to silver and records that it did. The authors state the limit themselves: the security property 'does not hold against collusion between the RA and TL operators', and the remedies they name are future work.
- 5BI-FIND-2026-1056 · A draft format for signed AI agent session records warns a signature does not vouch for what the record saysThe 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.
›In our briefings
The claim each of the newest briefings on SCITT would have you leave with.
Figuera's Sello protocol moves the record of an AI agent's actions from the agent to the services it calls. Each Sello-aware service seals a receipt body with HPKE to an owner key that the agent's JWS authorization token carries as a signed claim (owner_hpke_pk), signs the envelope with its own Ed25519 key and submits it to a public transparency log, where the owner finds it by the token's SHA-256 hash. In the paper's words, "The agent never holds any of the signing keys involved." The design assumes at least one service the agent calls is honest and uncompromised within the relevant time window: if every service colludes with the agent's operator, Figuera says, Sello cannot detect what they fabricate.
Sello has the services an AI agent calls sign its receipts; the agent can't fake them, but it can go around themBI-FIND-2026-2275
Revision -02 reverses the draft's rule on which agents count when a service admits a call. Revision -01 had the Serving Party check the principal and every actor in the delegation chain, which it said went beyond RFC 8693 token exchange, where earlier actors are informational only. Now the service checks the principal and the current actor, the agent making the call. Earlier actors named in nested act (actor) claims "are informational only, are recorded in actor_chain as audit history, and MUST NOT be used to decide admission," which the draft says Section 4.1 of RFC 8693 requires. The authors call the chain "attribution, not authority": the token issuer's asserted history, which is not complete and is no evidence that any hop held or narrowed authority. They concede that a Serving Party cannot refuse a call because an earlier actor was suspended; linked verification shows after the fact that the work continued. Admission on the whole actor history is an open issue, proposed for discussion at IETF 127 and not defined in the draft. Nor can the receipts show mechanically that authority narrowed at each hop: that would need a common way to represent scopes, another open issue, and the draft defines no mechanism for carrying authority across hops, pointing instead to delegation-chain mechanisms such as the OAuth identity-chaining draft.
Agent-audit draft now admits a call on the principal and the current agent, as RFC 8693 requiresBI-FIND-2026-2152
The draft puts the agent's actions in four new vCon event types: tool-call, tool-result, reasoning and system-event. Its compatibility section says vCon software that does not support the extension ignores those events and still processes the session's messages as usual, so "the agent trace is not processed". The authors call that compatible, and the "ai_agent" name is not required in the critical parameter. Those same events are where the draft asks for privacy care: tool-call inputs may contain credentials, API keys or file contents, and tool results may hold personal data. A redactor that does not support the extension cannot assess those fields and, the draft says, may reject a vCon that lists "ai_agent". The security section adds that malicious prompts kept in a record may execute if an AI system processes the record later.
A draft rebuilds agent session records on vCon, where software without the extension ignores the tool callsBI-FIND-2026-2135
›Cited in 30 briefings
Every briefing that discusses this term, newest first.
- Sello has the services an AI agent calls sign its receipts; the agent can't fake them, but it can go around them
Paper · · arXiv · Juan Figuera · Jun 2, 2026
- Agent-audit draft now admits a call on the principal and the current agent, as RFC 8693 requires
Draft · · IETF Datatracker · Shameer Thaha et al. · Oct 19, 2026
- A draft rebuilds agent session records on vCon, where software without the extension ignores the tool calls
Draft · · IETF Datatracker · Henk Birkholz et al. · Nov 2, 2026
- An agent-evidence draft no longer checks every actor in a delegation chain before a service admits a call
Draft · · IETF Datatracker · Shameer Thaha et al. · Oct 19, 2026
- A draft joining four agent records on one hash reports no valid joint run and now claims less about what agents did
Draft · · IETF Datatracker · Steven Mih et al. · Aug 16, 2026
- A SCITT draft tells verifiers not to guess how a hash was made; its interop proof is the authors' own record
Draft · · IETF Datatracker · Steven Mih et al. · Oct 26, 2026
- An agent audit draft says the agent is not trusted, yet lets the agent's operator run the audit store
Draft · · IETF Datatracker · Mirja Kühlewind et al. · Sep 7, 2026
- An agent can act with permission on a quarantined belief; a Glyphzero draft lets the action record show it
Draft · · IETF Datatracker · Karthik Rampalli, Glyphzero Labs · Jul 4, 2026
- AGTP's merchant draft has agents check the seller before paying, but the seller's own server runs the swap check
Draft · · IETF Datatracker · Chris Hood, Nomotic, Inc. · Oct 30, 2026
- An unadopted AGTP draft would let gateways refuse an agent's out-of-scope requests from its TLS certificate
Draft · · IETF Datatracker · Chris Hood, Nomotic, Inc. · Oct 30, 2026
- A draft has each company sign its own half when one company's agent acts at another's request, refusals included
Draft · · IETF Datatracker · Steven Mih, Action State Group · Oct 29, 2026
- A draft makes an AI system's refusals auditable in a SCITT log, but only for requests the system logged
Draft · · IETF Datatracker · Tokachi Kamimura, VeritasChain Standards Organization · Oct 24, 2026
- A 'human approved' field is not evidence, says Schrock's draft; only a bound, signed approval counts
Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Jul 3, 2026
- An agent audit log now names whom the agent acted for, but at low trust levels the agent may write that itself
Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Oct 26, 2026
- The SCITT API heading to RFC leaves who may register a statement to each operator
Draft · · IETF Datatracker · Henk Birkholz et al. · Jun 26, 2026
- RFC 9943 makes a log prove who signed a statement, not that it is true or that nothing was left out
Article · · IETF (RFC Editor) · Henk Birkholz et al. · Jun 30, 2026
- A transparency-log profile for agents exposes edited and dropped records; an action nobody logged leaves no gap
Draft · · IETF Datatracker · Pinar Emirdag et al. · Oct 12, 2026
- A draft records every AI agent verdict, refusals too, and won't let a dispatched action pass as confirmed
Draft · · IETF Datatracker · Steven Mih, Action State Group · Sep 27, 2026
- Bu's checklist names a verifier for each agent-protocol claim and checks 'the same agent as before' separately
Draft · · IETF Datatracker · Songbo Bu · Oct 14, 2026
- A post-quantum agent certificate draft warns that its own declared scopes are not proof of behaviour
Draft · · IETF Datatracker · Reuben Burls · Jul 21, 2026
- Schrock's receipts prove a named key approved one exact agent action, and the draft lists at length what they cannot prove
Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Sep 12, 2026
- Munoz's agent Permit is signed before the action runs, but the profile does not make the action wait for it
Draft · · IETF Datatracker · Christian Munoz, Keel API, Inc. · Jul 19, 2026
- Before an agent pays or changes data, Schrock's last check treats it as untrusted and keeps the record itself
Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Sep 26, 2026
- OpenA2A's protocol proves an agent holds a key and scores its behaviour, but never says who it works for
Draft · · IETF Datatracker · Abdel Fane, OpenA2A · Oct 6, 2026
- A draft ties what each AI agent may read to what it may do, and still can't take a grant back
Draft · · IETF Datatracker · Hamdy Atakora · Sep 21, 2026
- DIF's agent spec names every AI agent by a DID; the IETF's Agent Name Service switched to domain names
Article · · Decentralized Identity Foundation (KYA-OS) · Dylan Hobbs et al. · Jul 29, 2026
- The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted it
Draft · · IETF Datatracker · Scott Courtney et al. · Apr 13, 2026
- A valid delegation chain cannot replace required human approval or proof of key possession
Draft · · IETF Datatracker · Karthik Rampalli, Glyphzero Labs · Jul 6, 2026
- A proposed standard for agent authority advises against delegating it across company lines by default
Paper · · arXiv · Partha Madhira · May 12, 2026
- A draft format for signed AI agent session records warns a signature does not vouch for what the record says
Draft · · IETF Datatracker · Henk Birkholz et al. · Feb 25, 2026
Related terms
What appears alongside SCITT in the same briefings.
RFC 9943 SCITT architectureCertificate TransparencyCOSEAgent Action Capsule (SCITT)JSON Canonicalization SchemeRFC 9942 COSE ReceiptsAudit architecture for agent delegationCBOR
Was this explainer useful?
Share this explainer
Pass it on. Every claim has a permanent number, so what you share today still points at the same words next year.