MIAF pre-draft security review

[Zangarus]

Bug description

Findings are all related to the Margo (pre-draft) specification, not to any implementation / executable code. Therefore after discussions with the TWG Chairs, the security policy does not apply and there is no need to report the findings as a vulnerability according to the policy.

Findings come from a review of pre-draft, re-verified at f6c2305a28 (2026‑08‑27) after #194 and #25 merged and after #205 (RFC 9457 problem details) and #211 were opened.

ID	Severity	Title
01	High	Require an explicit origin allowlist for trustBundleUri
02	High	Restore rollback detection and add a staleness signal after spiffe_sequence removal
03	High	No replay protection: constrain TLS 1.3 early data on the Management Interface
04	High	Require integrity protection of cached trust material at rest
05	High	Desired state cannot pin a workload artifact by immutable digest
06	Medium	Define what a device MUST refuse in a Compose workload, and pin the Compose Specification version
07	High	MIAF sets no floor for key protection or SVID lifetime
08	Medium	Secure the proxy-to-backend hop in TLS-offload topologies, per RFC 9440 Section 4
09	Medium	Add verification guardrails for PKCS#1 v1.5 signatures and ECDSA nonce generation
10	Medium	Constrain Trust Bundle anchor entries: x5c count, JWK binding, and CA-ness
11	Medium	Require an authenticated time source or an anti-rollback store for certificate validation
12	Medium	Define a minimal MIAF audit event set
13	Medium	State that identity failure halts the management plane but does not stop running workloads
14	Medium	Bound certificate chain depth and presented-chain size
15	Medium	State the security cost of a shared wfm-id, and require per-instance key pairs
16	Medium	Define testable output constraints for MIS conformance
17	Medium	Bound Compose Archive entry count, uncompressed size and compression ratio
18	Medium	revision comparison is ambiguous: the tag mapping is lossy and SemVer precedence ignores build metadata
19	Medium	DeviceId permits . and .. path segments
20	Low	Bundle endpoint resilience: pin rotation, refresh jitter and bound, rate limiting
21	Low	TLS 1.2 fallback opt-in, and an unregistered problem type in the 403 example
22	Low	Enforce the /margo/ namespace at issuance, and a question about the removed body type discriminator
23	Low	State a hybrid key-exchange position and the chain-size implications of PQ signatures
24	Informational	Add a threat row for a compromised or malicious see-thru gateway
25	Low	Endpoint 404 wording contradicts the caller-scoping rule, and two navigation manifests are wrong
26	Strategic	Authenticated supply chain: signed desired state and attestation-backed issuance
27	Low	Identity concepts pages: late-binding cost, TLS-offload hop owner, bootstrap authentication in the diagram
28	Low	The inspection exemption asks operators to deviate from their own IEC 62443 zone-and-conduit policy

Detailed explanations to each finding are provided in the attached filed 00-consolidated-issue.md.

Proposed fix

Proposed solutions are incorporated in the explanation file 00-consolidated-issue.md.

Anything else (optional)

No response

[matlec]

Thanks again for this thorough review! I have addressed the MIAF-related findings in #227. Some of your findings belong to the in-progress MIAF SUPs, margo/specification-enhancements#84 (acquisition) and margo/specification-enhancements#95 (enrollment, renewal, and revocation), and I have addressed them in these SUPs. By the way, margo/specification-enhancements#95 proposes CRLs as a fourth revocation mechanism. If it is adopted, Trust Bundle rotation is no longer the only cryptographic revocation path (relevant to 02, 07, and 26).

ID	Resolution
01	Changed. A client that trusts public CAs rejects a discovered trustBundleUri on another origin, including after a redirect. An operator who serves the bundle elsewhere supplies trustBundleUri directly. The trustDomain check is now a MUST. The enrollment SUP applies the same rule to the enrollment endpoint.
02	Changed, in part. A client logs a spiffe_sequence regression, signals an old bundle, and caps spiffe_refresh_hint. It does not reject a bundle because of the sequence, so existing SPIFFE deployments stay conformant. The bundle is unsigned, so the sequence cannot detect a compromised MIS anyway.
03	Changed. No TLS 1.3 early data: a client does not send it, and an mTLS server does not accept it. New threat row.
04	Changed. SHOULD protect cached trust material at rest, plus a threat row. Not a MUST, because constrained devices may have no keystore.
05	Outside MIAF, still open (application description, desired state). SUP-01 (margo/specification-enhancements#68) rejected a separate digest field, so a fix needs to revisit that decision. margo/specification-enhancements#99 moves the revision and integrity rules for components into Margo Core.
06	Outside MIAF, still open (Compose). margo/specification-enhancements#99 plans a Compose deployment specification with a Compose Specification version statement and runtime requirements, which is the natural home for both halves. Non-root execution is tracked in #37.
07	In the SUPs. The acquisition SUP: products state how they protect their private key. The enrollment SUP: the MIS enforces an operator-configured maximum SVID lifetime. No fixed value, because the right lifetime depends on the offline window of the deployment.
08	Changed. The connection from the proxy to the backend MUST be protected against eavesdropping and modification (RFC 9440 §4). mTLS is preferred, not required, because some managed cloud load balancers cannot present a client certificate. New threat row.
09	Adopted in part. A verifier rejects RSA keys with an exponent below 65537. No rules for PKCS#1 v1.5 verification or ECDSA nonces: they describe library internals that no test can observe.
10	Changed, in part. A client ignores a bundle entry whose certificate is not a CA certificate. The other points are already in SPIFFE X.509-SVID §6. A bad entry is ignored, not the whole bundle, as SPIFFE requires.
11	Adopted in part. SHOULD use an authenticated time source (NTS) where available, and a skew limit of 5 minutes. The anti-rollback store is still open (device-requirements.md). No fail-closed rule, because it would stop devices at cold boot.
12	Changed. New sections on security event logging and non-repudiation.
13	Changed. An identity failure stops Management Interface calls, but not deployed workloads.
14	Changed. A verifier rejects a presented chain of more than six certificates, or larger than 64 KiB. Six covers the MIS deployment patterns, including nested SPIRE.
15	Changed, in part. The profile states the cost of a shared wfm-id. No requirement for a key pair for each instance: the enrollment SUP defers that pattern.
16	Changed. The MIS section lists what a test can check on issued SVIDs.
17	Outside MIAF, still open (Compose archive). margo/specification-enhancements#99 moves the archive safety rules into the Compose deployment specification.
18	Outside MIAF, still open (component revision). The precedence half is valid. The injectivity half does not hold: SemVer identifiers cannot contain _. The _ convention comes from SUP-00, The Helm Way (margo/specification-enhancements#67), and SUP-01 (margo/specification-enhancements#68). #143 covers the version of the ApplicationDescription, not the component revision.
19	Outside MIAF, still open (Management Interface). #225 rewrites exactly these patterns (renamed to targetName, at most two segments), but its patterns still accept . and ... The rule fits best into #225. Related: #191.
20	Changed. Refresh jitter, backoff, and rate limiting, plus a threat row. Pin rotation is already in the acquisition SUP §5.
21	Changed. TLS 1.2 is off by default and logged. The 403 prose names not-authorized. The specific wording stays in detail.
22	Changed, in part. The MIS does not issue a /margo/ path that no profile defines. The profile recognition rules already reject unknown paths at the verifier. Item 2 is outside MIAF and still open.
23	Changed. Hybrid key exchange is permitted, and the crypto-agility note states that PQ signatures and keys are much larger, and gives the service life of devices.
24	Changed. New threat row, close to your proposal.
25	Changed (items 1 and 4). Items 2 and 3 were already fixed on pre-draft.
26	Changed, in part. The framework states what MIAF does not take over from SPIFFE. The enrollment SUP answers how the MIS authenticates the CSR sender. Attestation and signed manifests stay on the roadmap.
27	No change. The concept pages are informative, and none of the items makes them wrong. The acquisition SUP §7 covers the cost of moving a client to another WFM.
28	Changed. The exemption stays, with informative guidance for a 62443 deviation and conduit controls that need no decryption. A TLS-offloading proxy is not a control point for the plant, because the operator often does not control it.

Still open outside MIAF / TL;DR

I did not address the following findings in my PR since they are not part of MIAF and might need broader discussion in the community.
I leave this issue open for these items.

• 05, 06, 17: application packaging and Compose
• 18: comparison of the component revision
• 19: dot segments in deviceId
• 22 (item 2): state what a server does with a request body of the wrong manifest type, now that the endpoint alone identifies the type

@ajcraig @phil-abb What's your opinion: should we split this issue into sub-issues, or keep it open and address the remaining findings in separate PRs that reference it?

[phil-abb]

@ajcraig @phil-abb What's your opinion: should we split this issue into sub-issues, or keep it open and address the remaining findings in separate PRs that reference it?

Yeah, it probably would be good to create a few different issues, with the items related to the same parts grouped.