Hermes Labs CISA CVSS Consistency Correction
A bounded audit by a Hermes Labs autonomous system found one CVSS score/vector mismatch in public enrichment data. The system stopped at the evidence boundary, a human approved the filing, and CISA republished the corrected score.
The correction
The CISA ADP entry for CVE-2026-14216 stored baseScore: 5.3 beside the vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:L. Under the FIRST CVSS 3.1 equations, that vector calculates to 6.5. The issue proposed two logically valid repairs: align the score to 6.5, or change the vector to one that calculates to 5.3.
CISA thanked the reporter and confirmed a corrected republish at 6.5. The official CVE corpus commit shows the numeric change from 5.3 to 6.5 while the vector and Medium category remain unchanged. A later Vulnrichment data-update commit synchronized the refreshed record. This was a record-consistency correction, not discovery of a new vulnerability or a change in CISA's severity category.
From mandate to bounded audit
On September 2, 2026, the owner's mandate opened and closed with:
“MISSION: CHANGE THE WORLD-STATE”
“Change the world-state. Leave independently observable evidence that Hermes Labs caused the change.”
A later owner instruction authorized a coordinator to take obvious bounded next actions without broadening scope. The Hermes Labs system chose the CISA audit because it offered a public correction path, deterministic validation, bounded attention cost, and an independently controlled receipt. The coordinator then launched only local and read-only execution: investigation, tests, and correction preparation. External publication remained parked.
The audit fixed the Vulnrichment develop snapshot at b4481a9ea6cbb1e36ba9a746ef29de45b95567f8, selected a 14-day window, capped supporting sources, defined deterministic CVSS, CWE, CPE, and SSVC checks, excluded duplicates and open issues, and allowed a 30-day expansion only if the first pass returned no findings.
What the run checked
The session-reported full pass examined 178,776 repository JSON records and selected 5,324 eligible recent CISA ADP records. Within that set it checked 1,060 CVSS 3.x metrics, 1,634 CWE assignments, and 5,324 SSVC blocks. Exactly one deterministic mismatch survived.
Those are distinct denominators. The run did not claim that every eligible record contained a CVSS vector, and one finding is not an error-rate estimate. These aggregates come from the publisher-held execution record and its hashed result, not from a new replay performed for this article.
The nonzero result activated the stop rule: the system did not widen to 30 days. With only one finding, it also abandoned the planned five-filed/five-held-out experiment rather than manufacture a larger sample to satisfy the design.
Correct the claim before the record
An initial draft inferred from repository history that CISA had changed the vector's availability impact from A:N to A:L. A subsequent history check showed the record first appeared already inconsistent. The system withdrew the inference before publication and narrowed the report to what the evidence proved.
That correction changed the issue from a story about how the mismatch arose into a precise consistency report. It preserved both valid resolution paths and avoided attributing a causal edit that the public history did not show.
The public-action boundary
The system prepared the correction but stopped before creating the GitHub issue. At , Roli Bosch authorized the exact target and account: “AUTHORIZED. Post the prepared issue to cisagov/vulnrichment as roli-lpci.” The same instruction accepted the single reproduced inconsistency, prohibited expansion for the holdout design, and required a narrowly technical report without Hermes promotion.
Before posting, the system rechecked the account, target record, duplicate-issue state, and CVSS result. Bosch's account is the public issue author. Hermes Labs' role is established by the publisher-held execution record summarized here; the issue itself does not claim Hermes attribution or CISA endorsement.
Public timeline
Open the public chronologyIssue, implementation, and merge dates
- · Issued a broad world-state mandate with an independently observable evidence requirement.
- · Authorized a coordinator to take obvious bounded next actions while prohibiting scope expansion.
- · Selected the CISA Vulnrichment audit from at most three candidates using deterministic validation and public-correction criteria.
- · Continued the selected audit for local and read-only work while keeping external mutations parked.
- · Froze a 14-day snapshot audit, deterministic validators, exclusions, and a conditional null-yield expansion rule.
- · Found one score/vector mismatch in CVE-2026-14216 and reproduced the 6.5 calculation.
- · Withdrew an unsupported history inference and narrowed the proposed correction language.
- · Authorized the exact filing target and account and directed the system not to expand the audit for an artificial holdout.
- · Opened cisagov/vulnrichment issue #333 with the reproduced inconsistency. Primary source.
- · The official CVE corpus recorded the numeric correction from 5.3 to 6.5. Primary source.
- · Thanked the reporter, confirmed the republished 6.5 score, and closed the issue. Primary source.
- · Vulnrichment synchronized the refreshed upstream record in its data-update commit. Primary source.
Inspect and reuse the method
The machine-readable case record publishes stable event IDs, UTC timestamps, actor/action attribution, evidence classes, snapshot and aggregate fields, and primary-source citations. Publisher-held events are summaries of the execution record. Raw transcripts, private URLs, local paths, credentials, unrelated session work, and hidden reasoning are excluded.
The Hermes Labs CVSS Consistency repository at the reviewed publication commit preserves the recovered original scanner separately and provides a reviewed public-record checker for reuse. Its fixture replay is a post-event implementation, not a rerun of the original 178,776-record audit.
The practical pattern is portable: freeze the source state, define deterministic checks and stop rules before scanning, surface only reproducible anomalies, correct unsupported provenance claims, verify freshness, and require exact authorization at the external-action boundary. This case documents one governed path to a useful correction; it does not establish a general safety property for autonomous systems.
Case ID: cisa-vulnrichment-333 · Status: acknowledged-and-corrected · All case studies