Skip to content

fix(monitor): treat PERSIST-002 and MULTI-001 as high-FP single-rule verdicts#33

Merged
Ap6pack merged 1 commit into
mainfrom
claude/project-audit-modernize-lda8id
Jul 17, 2026
Merged

fix(monitor): treat PERSIST-002 and MULTI-001 as high-FP single-rule verdicts#33
Ap6pack merged 1 commit into
mainfrom
claude/project-audit-modernize-lda8id

Conversation

@Ap6pack

@Ap6pack Ap6pack commented Jul 17, 2026

Copy link
Copy Markdown
Owner

What / why

Works the top item of #30: auditing the 82 MALICIOUS verdicts that still rested
on a single unverified rule. Running those rules against benign inputs (the same
method that caught CMD-001/ENV-001) surfaced two more that over-flag:

  • MALWAR-PERSIST-002 matches self-referential file ops on
    SKILL.md / CLAUDE.md / .claude/ — including reads (cat SKILL.md | grep)
    and legitimate skill-authoring tools that write a CLAUDE.md.
  • MALWAR-MULTI-001 matches benign prose about quiet/deferred execution
    ("applies the patch without showing the full diff").

Both join HIGH_FP_RULES, so a single-rule MALICIOUS verdict on them is
verified-or-downgraded, not auto-convicted.

MALWAR-PERSIST-001 (cron / systemd / .bashrc) did not over-flag in
testing and deliberately stays a confident single-signal rule.

Impact

  • Recall preserved: rules still fire on the malicious fixtures; corroborated
    (2+ rule) verdicts unaffected.
  • Shrinks the unverified single-rule MALICIOUS tier from 82 → 39 on the
    current snapshot (43 PERSIST-002/MULTI-001 cases now verify-or-downgrade).
  • Full unit suite: 1440 passing; new tests assert the two rules are fragile
    and PERSIST-001 is not.

Note

Numbers update on the next full re-scan (genuinely-malicious cases get
LLM-confirmed back; false positives downgrade). Does not by itself close #30
the remaining 39 sit on tighter rules that didn't over-flag, plus the other
checklist items (detector novelty logic, human-approval gate, dry-run mode).

🤖 Generated with Claude Code


Generated by Claude Code

…verdicts

Auditing the residual single-rule MALICIOUS tier (issue #30) surfaced two more
rules that over-flag legitimate content when they alone drive a conviction,
confirmed by running them against benign inputs:

- MALWAR-PERSIST-002 matches self-referential SKILL.md/CLAUDE.md/.claude file
  ops including reads ('cat SKILL.md | grep') and legit skill-authoring tools
  that write a CLAUDE.md.
- MALWAR-MULTI-001 matches benign prose about quiet/deferred execution
  ('applies the patch without showing the full diff').

Both join HIGH_FP_RULES, so a single-rule MALICIOUS verdict on them is
verified-or-downgraded rather than auto-convicted. MALWAR-PERSIST-001
(cron/systemd/.bashrc) did NOT over-flag in testing and stays a confident
single-signal rule. Recall preserved: the rules still fire on the malicious
fixtures; corroborated (2+ rule) verdicts are unaffected.

Shrinks the unverified single-rule MALICIOUS tier from 82 to 39 on the current
snapshot. Full unit suite (1440) passes.

Co-Authored-By: Claude <noreply@anthropic.com>
@Ap6pack
Ap6pack merged commit 1f590c6 into main Jul 17, 2026
5 checks passed
@Ap6pack
Ap6pack deleted the claude/project-audit-modernize-lda8id branch July 17, 2026 08:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Re-validate detect → verify → publish path before re-enabling auto-publish

2 participants