chore(llm-security): STATE — security-fix track plan (F-1/F-2/F-3 must-fix) over multiple sessions

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3s6WnubSSrFjAQTLQdVbG
This commit is contained in:
Kjell Tore Guttormsen 2026-06-20 10:09:11 +02:00
commit a96e3b947f

View file

@ -1,37 +1,40 @@
# STATE — llm-security
**Active focus:** TRG/SIG/AST-gap scanners (TRG/SIG/AST) — DONE & pushed (Steps 17).
Release v7.8.0 (Step 8) DEFERRED by operator. **The next session is NOT TRG/SIG/AST** — it is
the security-fix brief (see PREREQUISITE) which must run before we resume/release this track.
**Active focus:** Security-fix track (from `docs/security-fix-brief-2026-06-20.md`). Critical review
is DONE — brief verified sound. Implement the must-fixes over MULTIPLE sessions (context-limited).
TRG/SIG/AST (TRG/SIG/AST) shipped Steps 17; v7.8.0 release (Step 8) stays DEFERRED behind these fixes.
## PREREQUISITE — do BEFORE continuing TRG/SIG/AST / releasing v7.8.0
- Run the security-fix session: **`docs/security-fix-brief-2026-06-20.md`**
(F-1 RCE + F-2/F-3 shell/path injection sinks; from the marketplace-wide review
`docs/review-2026-06-20.md`). Operator instruction (2026-06-20): do NOT interleave it with
TRG/SIG/AST work; it gates the v7.8.0 release — don't ship a version with F-1/F-2/F-3 open.
## Critical-review verdict (DONE — do NOT re-derive)
- **F-1 / F-2 / F-3 = MUST FIX.** Verified real, correctly diagnosed, fixes are local + mechanical.
- **F-5 / F-6 = cheap cleanup.** `git rm -- ./--json .orphaned_at` — both confirmed present in repo root.
- **F-4 = optional** (LOW, by-design MCP spawn; a confirm-gate is a judgment call).
- No false positives / over-scoping in the must-layer. One gap to add: F-2's prefix containment does
not stop a symlink-escape — note it, but the prefix check is the must-do part.
## What landed this session (pushed to Forgejo, main @ d19abf0)
- **TRG** `scanners/trigger-scanner.mjs` — shadow/baiting/broad triggers; decode-pipeline on
descriptions. + wiring (orchestrator/policy/4 OWASP maps). Commits 54115a9, 75aeeee. Test 16/0.
- **SIG** `scanners/signature-scanner.mjs` + `knowledge/signatures.json` (webshell/reverse_shell/
cryptominer/hacktool, matched over normalizeForScan/foldHomoglyphs/rot13 — catches obfuscated
malware). + wiring. Commits bac6f6f, 76e3543. Test 12/0.
- **AST** `scanners/ast-taint-scanner.mjs` + `scanners/lib/py-ast-taint.py` (PARSE-ONLY ast.parse,
scope-aware var taint, 5s timeout, graceful skip when python3 absent). + wiring.
Commits e497bb7, b50313e. Test 9/0. Parse-only proven (sentinel: no SENTINEL written).
- Docs/packaging `d19abf0` — scanner-reference rows + count bumps, output.mjs JSDoc prefixes,
orchestrator header (→14), `package.json` files[] += `knowledge/`.
- Full suite **1859/0**. Orchestrator runs **14** scanners (trg/sig/ast keys present). Zero-dep holds.
## Multi-session plan (TDD each: repro → red → green; full `node --test` green; gitleaks clean)
- **Session A — F-1 (CRITICAL) + cleanup.** Convert the `git()` helper in
`scanners/git-forensics.mjs` (`:59-66`, currently `execSync(\`git ${cmd}\`)`) to
`spawnSync('git', [...tokens], {cwd})` — one change fixes every call site (`:202/:211/:223/:231/
:549/:564` already pass discrete tokens). Add a fixture repo with a hostile filename; assert the
scan completes and the injected side-effect never runs (write failing repro FIRST). Keep existing
git-forensics tests green (glob `commands/*.md` must still match via git, not shell). Fold in
F-5/F-6 `git rm`. Commit + push.
- **Session B — F-2 + F-3 (both HIGH).**
- F-2 `scanners/auto-cleaner.mjs` (sink ~`:776`, write ~`:878-881`): before write assert
`absPath === targetPath || absPath.startsWith(targetPath + sep)`; else skip + report. Test: a
finding `file: "../escape.txt"` is refused, not written.
- F-3 `scanners/lib/supply-chain-data.mjs:221-227` (`execSafe` = `execSync`) reached via
`hooks/scripts/pre-install-supply-chain.mjs:254`: switch to `spawnSync('npm',['view',spec,'--json'])`
OR validate spec `^[@/A-Za-z0-9._-]+$` first. Test: spec `foo;touch X` must not execute `touch`.
- **Session C — release.** Then Step 8 (v7.8.0) per
`docs/plans/trekplan-2026-06-20-TRG/SIG/AST-gap-analysis.md`. Optionally F-4 confirm-gate.
## Step 8 (deferred) — release v7.8.0 when ready, AFTER the security fix
- Bump 7.7.2 → 7.8.0 in `.claude-plugin/plugin.json`, `package.json`, README badge (`README.md:9`),
CHANGELOG `[7.8.0]`, `docs/version-history.md`, CLAUDE.md header. Verify w/ version-consistency grep.
Spec: Step 8 of `docs/plans/trekplan-2026-06-20-TRG/SIG/AST-gap-analysis.md`.
## COLD START (next session = Session A)
1. `pwd` + `git status` (expect clean). Read `docs/security-fix-brief-2026-06-20.md` + this STATE.
2. Start Session A (F-1) TDD — write the failing repro FIRST, then convert `git()`.
## Known pre-existing doc drift (out of this plan's scope — flagged, NOT fixed)
- `docs/scanner-reference.md` orchestrated count/list never counted `workflow` (true array = 14, docs
now say 13). Per the plan I bumped each count by +3; reconciling the workflow omission needs its own scope.
## Continuity
- STATE.md canonical + committed (never gitignored). trekexecute progress scratch deleted (run done;
git history is the durable record). origin = Forgejo `open/llm-security` (never GitHub).
## Continuity notes
- Brief's "disclosure hold" gate is MOOT: review `738770e` + brief `fea943b` are already on
`origin/main`. Push the F-1 fix normally; no special bundling needed.
- origin = Forgejo `open/llm-security` (never GitHub). Push window: weekdays 20:0023:00; weekends/
holidays anytime. TRG/SIG/AST scanners (14 total, suite 1859/0) already shipped + pushed.