01Finding and impact
A derivative-write path in the private research-recording tool could create caller-influenced PROVEN/RESOLVED records without the shared validation used by ordinary writes. The demonstrated defect was omission of required falsifier metadata: a record of what would contradict the claim.
A downstream consumer could treat such a record as more authoritative than the tool's checks justified. The demonstrated scope is internal evidence integrity, not an established remote authentication break, cryptographic forgery or third-party compromise.
02Corrective change
The recorded repair applies shared validation to both result branches and changes attribution to distinguish a caller's claimed evaluation from an executed measurement.
The gate validates evidence structure, not authenticated measurement provenance. A later false claim still entered the record store and was detected and corrected through source and artifact checks. This repair does not establish complete prevention of false evidence or prompt injection.
03Recorded comparisons
A later driver used the existing local MCP interface in separate copies of the vulnerable source, repair and pinned later source. It appended the same synthetic fixture and recorded a successful-result derivative. The measurement-like fixture text was synthetic input, not an actual hardware measurement.
| Source copy | Recorded state | Falsifier metadata |
|---|---|---|
| Pre-repair | PROVEN/RESOLVED | Missing |
| Repair | PROVEN/RESOLVED | Present |
| Pinned later source | PROVEN/RESOLVED | Present |
All three server runs recorded exit status 0 and successful initialization. The direct comparison checks the successful-result metadata branch, which the existing fixture did not explicitly assert.
The current record-validation suite passed 45/45 checks on the pinned later source. On unchanged pre-repair implementation with the same current fixtures, 38 passed and 7 failed. Six failures were adversarial assertions; the seventh concerned a historical-ledger data expectation. It is not counted as a security repair.
04Source identity and evidence
- Pre-repair
- 8a1bfcfd5ebac049310f60fa5b3a08be5693ccac
- Repair
- c25c9d27e31ba92b293572bf7af76ee7ceaf50e2
- Later revision
- c109f4e439aa52b50ce6fd2183d04a8e7202ee5a
The documented hardening record is dated 24 June 2026. The later validation packet was prepared on 31 October 2026 UTC. Neither date is asserted as the first discovery date. The full affected-version range has not been established.
Download the observation summary (JSON), its method and source-availability note, and SHA-256 checksums. The summary is a curated extraction of recorded results. The source and original logs remain private; these downloads are not a complete public reproducer. Hashes identify bytes, not independent validation.
Work directed by Nayte Wilson, with AI assistance. Both advisories document one hardening effort in software operated by MacBrains Labs. The public ANE dataset and its DOI cover separate measurement summaries.