Unsigned claims can reach trusted status in Mareforma 0.3.3
A flaw in the Python library's convergence logic lets attackers without keys promote forged claims to REPLICATED, the status downstream consumers treat as multi-party endorsement.
By The Agentic Times ·

A critical integrity flaw in Mareforma 0.3.3, a Python library for recording and cross-checking claims in a shared graph, lets an attacker with no cryptographic key write unsigned claims that are promoted to REPLICATED status. The issue was published as GHSA-8776-qf3h-26r3, tracked as CVE-2026-71535, on 19 November 2026. According to the advisory, the forged rows are returned by the library's default trust query and are indistinguishable from legitimately converged claims.
Mareforma treats REPLICATED as a signal that two independent parties agree on the same anchor. The advisory says the function that decides whether two claims agree, _maybe_update_replicated_unlocked() in mareforma/db/core.py, selects peer claims with a SQL query that filters on status, resolution state, transparency logging and artefact hash. It contains no check on the signature_bundle field and no check that the claim came from an enrolled validator key. Once peer rows are collected, the advisory says, all matching claim IDs are promoted to REPLICATED with no further validation.
Three defaults make the conditions easy to meet, according to the advisory. First, transparency logging to Rekor is disabled by default, and when it is disabled the library sets the transparency_logged flag to 1 immediately after insert, so every claim becomes eligible for convergence as soon as it is written. Second, require_signed defaults to False. Third, if a caller opens the graph with a key path that does not exist, the library silently sets the signer to None rather than raising an error, which enables unsigned claim insertion. The advisory also notes that assert_claim() performs no signature gate and only validates DOI resolution on the supports list before passing attacker-controlled fields to the database layer.
The effect is that the default query path filters out only PRELIMINARY claims from generators that are not enrolled. Claims at REPLICATED status are returned unconditionally. The advisory says this allows an attacker to inject false trust signals into any downstream consumer that relies on REPLICATED as a multi-party endorsement guarantee.
The advisory argues the omission looks accidental rather than deliberate. A helper function, _extract_signature_bundle_keyid(), already exists in the same file and can extract a key ID from a signature bundle, but it is never called on the convergence path.
A proof of concept is included in the advisory. It installs Mareforma 0.3.3 in a virtual environment, has a legitimate root agent bootstrap a key and create an established anchor, has an honest peer assert a signed preliminary claim supporting that anchor, then has an attacker open the same graph with a non-existent key path to enter unsigned mode. The advisory also provides a Dockerfile so the demonstration can be run with no network access.
The affected package is mareforma on PyPI. The advisory lists the full data flow from the command line interface, which explicitly permits unsigned graph mode, through to the database update that writes REPLICATED status. No fixed version is stated in the advisory text.
Sources
- github.comhttps://github.com/advisories/GHSA-8776-qf3h-26r3
