Skip to content

PI079: the 40-character key-blob minimum is a choice not a measurement; no independent true-positive or false-positive evidence exists #180

Closed

Description

@hermanngeorge15

PI079 persistence-command-with-payload (MEDIUM) fires on one line carrying a literal SSH public key of a named type that appends it to authorized_keys with a redirect or an append-mode tee. It shipped on its written criterion (fires through the prose pass on a blind-written corpus payload, silent on every clean specimen), with these honest qualifications recorded at ship time:

  1. It adds no recall. PI077 already counts the one payload it reaches, so the pin was byte-identical; and it reaches none of the 12 held-out payloads.
  2. The 40-character minimum on the key blob is a choice, made so the elided form tutorials print (AAAAB3NzaC1yc2E...) stays silent. It is not derived from a measurement.
  3. No independent evidence either way. Zero lines carrying a literal key blob in the 26,407 files of the frozen GATE-03 inputs (so no false positives and no true positives observed), and the only key-blob lines on the author's machine were in that phase's own agent transcripts.
  4. MEDIUM was a judgment call under PATTERNS.md rule 3: a container build line appending a real public key, and an infrastructure runbook adding a named engineer's key, are recallable benign documents with the same line shape. It is one YAML field to flip.

Revisit when: a real document (benign or hostile) trips it, which would turn the threshold and the grade into measurements.


Filed at the close-out of Phase 5 (CAT-03, #35), plan 05-07 Task 2, from .planning/phases/05-persistence-lifecycle-hijack-cat-03-35/deferred-items.md. Nothing here is fixed in v0.2.0.

Activity

  1. added this to the v0.3.0 milestone on Oct 9, 2026
  2. BasilZafar11 commented on Oct 22, 2026

    @BasilZafar11
    Contributor

    Opened #200 addressing #180 with independent hostile and benign specimens, reproducible threshold measurements, and updated remediation guidance. The evidence supports retaining MEDIUM severity; calibration limits are documented. Includes regression tests and a regenerated baseline.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P2Normal prioritypatternsPattern library work

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions