fix(llm-security): commons-loader - drop policy-driven root, ship path, fix cache [skip-docs]

Advisor review on the prior commit (69cad7c) caught a real detection-kill
vulnerability before push: reading `commons.root` from the SCANNED
TARGET's .llm-security/policy.json let a hostile cloned repo redirect
llm-security's own detection corpus to an attacker-supplied (empty)
one, with graceful-empty fallback making the substitution silent — a
substitutive override, unlike sig.custom_rules_path's additive one.
Dropped the policy import entirely; commons location is this plugin's
own concern, resolved only from __dirname or an explicit test/dev
override, never from policy or the scan target.

Also fixed two issues the review surfaced:
- Default vendor path was repo-root `shared/`, which package.json's
  `files` allowlist (bin/, scanners/, knowledge/) would never publish —
  moved under scanners/commons/, inside the directory that actually
  ships. Same defect class as 2fe2915 (green dev checkout, empty
  detection tables once installed).
- Cache keyed success/failure together, so the first caller's
  `fallback` shape (e.g. []) leaked to a second caller expecting a
  different shape ({}) on the same missing artifact. Cache now stores
  a load-failed sentinel and returns each caller's own fallback.
  Loaded artifacts are also deep-frozen, since the cache hands out one
  shared object by reference to every caller.

New/changed tests cover all four: a simulated hostile-target policy
file is ignored, the failure-cache no longer cross-contaminates
fallback shapes, and mutating a loaded artifact throws.

Golden baseline unchanged; full suite 2063/2063.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QAYkRaBXT6tmWXTQAi1ZBg
This commit is contained in:
Kjell Tore Guttormsen 2026-08-09 14:11:56 +02:00
commit b0de0ca6d8
2 changed files with 92 additions and 71 deletions

View file

@ -7,51 +7,62 @@
// hardcoded tables to this loader in Phase 5 step 4, table-by-table, behind // hardcoded tables to this loader in Phase 5 step 4, table-by-table, behind
// the golden gate. // the golden gate.
// //
// Modeled on signature-scanner.mjs's loadRules()/loadCustomRules() pair: // Modeled on signature-scanner.mjs's loadRules(): hooks run per-tool-call in
// hooks run per-tool-call in fresh zero-dep processes, so resolution must be // fresh zero-dep processes, so resolution must be a fast synchronous read of
// a fast synchronous read of the vendored copy, never network. Cached once // the vendored copy, never network. Cached once per process. A
// per process. A missing/unvendored commons dir degrades to the caller's // missing/unvendored commons dir degrades to the caller's fallback rather
// fallback rather than crashing a hook — the same graceful-empty contract // than crashing a hook — the same graceful-empty contract loadRules() uses
// loadRules() uses for a missing knowledge/signatures.json. // for a missing knowledge/signatures.json.
//
// Deliberately NOT policy-driven, unlike loadCustomRules()'s
// sig.custom_rules_path: this plugin scans untrusted cloned repos, and
// `.llm-security/policy.json` lives in the *scanned target*. A
// custom_rules_path override can only ever add findings a hostile target
// supplies; a commons-root override would let that target *replace* this
// plugin's own detection corpus wholesale, with graceful-empty fallback
// making the substitution silent. The commons location is this plugin's own
// concern, not the scan target's — so the only override is the explicit
// `commonsRoot` option (tests, a local commons dev checkout), resolved from
// this file's own location, never from policy or the scan target.
// //
// Zero external dependencies — Node.js builtins only. // Zero external dependencies — Node.js builtins only.
import { readFileSync } from 'node:fs'; import { readFileSync } from 'node:fs';
import { join, dirname, isAbsolute, resolve } from 'node:path'; import { join, dirname } from 'node:path';
import { fileURLToPath } from 'node:url'; import { fileURLToPath } from 'node:url';
import { getPolicyValue } from './policy-loader.mjs';
const __dirname = dirname(fileURLToPath(import.meta.url)); const __dirname = dirname(fileURLToPath(import.meta.url));
// Default vendored location: llm-security-commons is pulled in as a // Default vendored location: llm-security-commons is pulled in as a
// pull-only subtree at the repo root, mirroring the portfolio-optimiser // pull-only subtree under scanners/, which is the only source directory
// family's `shared/` convention (docs/commons-extraction-plan.local.md). // `package.json`'s `files` allowlist ships (bin/, scanners/, knowledge/ —
const DEFAULT_COMMONS_ROOT = join(__dirname, '..', '..', 'shared'); // scripts/ and tests/ do not publish, and neither would a repo-root dir).
const DEFAULT_COMMONS_ROOT = join(__dirname, '..', 'commons');
// Cached, parsed artifacts, keyed by resolved absolute file path. // Cached parsed artifacts, keyed by resolved absolute file path. A failed
// load caches this sentinel (not the caller's fallback value) so that two
// callers requesting the same missing artifact with different fallback
// shapes (`[]` vs `{}`) each still get their own fallback back, rather than
// the first caller's shape leaking to the second.
const _cache = new Map(); const _cache = new Map();
const LOAD_FAILED = Symbol('commons-loader:load-failed');
/** /** Recursively freeze a parsed JSON value so callers cannot mutate a shared cached table. */
* Resolve the commons root directory. function deepFreeze(value) {
* Precedence: explicit `commonsRoot` option > `commons.root` policy value if (value !== null && typeof value === 'object' && !Object.isFrozen(value)) {
* (relative paths resolve against `targetPath`) > the default vendored path. Object.freeze(value);
* @param {string|undefined} targetPath for (const key of Object.keys(value)) deepFreeze(value[key]);
* @param {string|undefined} commonsRoot
* @returns {string}
*/
function resolveCommonsRoot(targetPath, commonsRoot) {
if (commonsRoot) return commonsRoot;
const policyRoot = getPolicyValue('commons', 'root', null, targetPath);
if (policyRoot && typeof policyRoot === 'string') {
return isAbsolute(policyRoot) ? policyRoot : resolve(targetPath || process.cwd(), policyRoot);
} }
return DEFAULT_COMMONS_ROOT; return value;
} }
/** /**
* Load and parse one commons JSON artifact. * Load and parse one commons JSON artifact.
* Graceful fallback on any read/parse error an unvendored, missing, or * Graceful fallback on any read/parse error an unvendored, missing, or
* invalid commons artifact must never crash a hook or scanner. * invalid commons artifact must never crash a hook or scanner. The returned
* value is deep-frozen: every caller shares the same cached object, so
* mutating it would silently leak across callers/targets; freezing makes
* that throw instead.
* *
* @param {string} artifactPath - relative path under the commons root, * @param {string} artifactPath - relative path under the commons root,
* without the `.json` extension (e.g. `'lexicon/injection-lexicon'`). * without the `.json` extension (e.g. `'lexicon/injection-lexicon'`).
@ -59,25 +70,27 @@ function resolveCommonsRoot(targetPath, commonsRoot) {
* @param {*} [opts.fallback] - value returned on any load/parse failure * @param {*} [opts.fallback] - value returned on any load/parse failure
* (default: `null`). Callers pick the shape-appropriate empty value * (default: `null`). Callers pick the shape-appropriate empty value
* (`[]`, `{}`, ...) the same way loadRules() falls back to `[]`. * (`[]`, `{}`, ...) the same way loadRules() falls back to `[]`.
* @param {string} [opts.commonsRoot] - explicit commons root, overriding * @param {string} [opts.commonsRoot] - explicit commons root, overriding the
* policy and the default (for tests and one-off callers). * default vendored path (for tests and a local commons dev checkout).
* @param {string} [opts.targetPath] - scan root used to resolve a * @returns {*} Parsed, frozen JSON, or `opts.fallback` on failure.
* policy-relative `commons.root` and to locate `.llm-security/policy.json`.
* @returns {*} Parsed JSON, or `opts.fallback` on failure.
*/ */
export function loadArtifact(artifactPath, opts = {}) { export function loadArtifact(artifactPath, opts = {}) {
const { fallback = null, commonsRoot, targetPath } = opts; const { fallback = null, commonsRoot } = opts;
const root = resolveCommonsRoot(targetPath, commonsRoot); const root = commonsRoot || DEFAULT_COMMONS_ROOT;
const filePath = join(root, `${artifactPath}.json`); const filePath = join(root, `${artifactPath}.json`);
if (_cache.has(filePath)) return _cache.get(filePath); if (_cache.has(filePath)) {
const cached = _cache.get(filePath);
return cached === LOAD_FAILED ? fallback : cached;
}
let result; let result;
try { try {
const raw = readFileSync(filePath, 'utf8'); const raw = readFileSync(filePath, 'utf8');
result = JSON.parse(raw); result = deepFreeze(JSON.parse(raw));
} catch { } catch {
result = fallback; // graceful: unvendored/missing/invalid -> caller's empty shape _cache.set(filePath, LOAD_FAILED); // graceful: unvendored/missing/invalid -> caller's fallback
return fallback;
} }
_cache.set(filePath, result); _cache.set(filePath, result);

View file

@ -12,12 +12,9 @@ import { writeFileSync, mkdirSync, rmSync } from 'node:fs';
import { join } from 'node:path'; import { join } from 'node:path';
import { tmpdir } from 'node:os'; import { tmpdir } from 'node:os';
import { loadArtifact, _resetCacheForTest } from '../../scanners/lib/commons-loader.mjs'; import { loadArtifact, _resetCacheForTest } from '../../scanners/lib/commons-loader.mjs';
import { _resetCacheForTest as _resetPolicyCacheForTest } from '../../scanners/lib/policy-loader.mjs';
const TEST_ROOT = join(tmpdir(), `llm-security-commons-loader-test-${Date.now()}`); const TEST_ROOT = join(tmpdir(), `llm-security-commons-loader-test-${Date.now()}`);
const FIXTURE_COMMONS_ROOT = join(TEST_ROOT, 'fixture-commons'); const FIXTURE_COMMONS_ROOT = join(TEST_ROOT, 'fixture-commons');
const POLICY_DIR = join(TEST_ROOT, '.llm-security');
const POLICY_FILE = join(POLICY_DIR, 'policy.json');
function writeArtifact(relPath, data) { function writeArtifact(relPath, data) {
const filePath = join(FIXTURE_COMMONS_ROOT, `${relPath}.json`); const filePath = join(FIXTURE_COMMONS_ROOT, `${relPath}.json`);
@ -28,13 +25,11 @@ function writeArtifact(relPath, data) {
describe('commons-loader', () => { describe('commons-loader', () => {
beforeEach(() => { beforeEach(() => {
_resetCacheForTest(); _resetCacheForTest();
_resetPolicyCacheForTest();
mkdirSync(FIXTURE_COMMONS_ROOT, { recursive: true }); mkdirSync(FIXTURE_COMMONS_ROOT, { recursive: true });
}); });
afterEach(() => { afterEach(() => {
_resetCacheForTest(); _resetCacheForTest();
_resetPolicyCacheForTest();
try { rmSync(TEST_ROOT, { recursive: true }); } catch {} try { rmSync(TEST_ROOT, { recursive: true }); } catch {}
}); });
@ -57,9 +52,11 @@ describe('commons-loader', () => {
assert.deepEqual(data, {}); assert.deepEqual(data, {});
}); });
it('degrades gracefully when commons has not been vendored yet (no override, no policy)', () => { it('degrades gracefully when commons has not been vendored yet (no commonsRoot override)', () => {
// Phase 4 hasn't run in this repo checkout: the default `shared/` vendor // Phase 4 hasn't run in this repo checkout: the default scanners/commons
// path does not exist. The loader must not throw. // vendor path does not exist. The loader must not throw, and — since the
// commons location is never policy- or target-derived — this cannot be
// influenced by an on-disk .llm-security/policy.json either.
const data = loadArtifact('lexicon/injection-lexicon', { fallback: 'EMPTY' }); const data = loadArtifact('lexicon/injection-lexicon', { fallback: 'EMPTY' });
assert.equal(data, 'EMPTY'); assert.equal(data, 'EMPTY');
}); });
@ -73,7 +70,7 @@ describe('commons-loader', () => {
assert.equal(second.entropyFloor, 4.5); // original value, not the mutation assert.equal(second.entropyFloor, 4.5); // original value, not the mutation
}); });
it('caches a failed load too, so a fixed-but-unread file still returns the cached fallback', () => { it('caches a failed load too, so a fixed-but-unread file still returns the fallback', () => {
const first = loadArtifact('signatures/malware-signatures', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: [] }); const first = loadArtifact('signatures/malware-signatures', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: [] });
writeArtifact('signatures/malware-signatures', { rules: [{ id: 'late-arrival' }] }); writeArtifact('signatures/malware-signatures', { rules: [{ id: 'late-arrival' }] });
const second = loadArtifact('signatures/malware-signatures', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: [] }); const second = loadArtifact('signatures/malware-signatures', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: [] });
@ -81,36 +78,47 @@ describe('commons-loader', () => {
assert.deepEqual(second, []); assert.deepEqual(second, []);
}); });
it('resolves commonsRoot from the commons.root policy value when no explicit override is given', () => { it('a cached failure does not leak one caller\'s fallback shape to another', () => {
mkdirSync(POLICY_DIR, { recursive: true }); const arrayFallback = loadArtifact('missing/artifact', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: [] });
writeFileSync(POLICY_FILE, JSON.stringify({ commons: { root: FIXTURE_COMMONS_ROOT } })); const objectFallback = loadArtifact('missing/artifact', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: {} });
writeArtifact('codepoints/carriers', { zeroWidth: [''] }); assert.deepEqual(arrayFallback, []);
const data = loadArtifact('codepoints/carriers', { targetPath: TEST_ROOT, fallback: null }); assert.deepEqual(objectFallback, {}); // not [] from the first caller's cached fallback
assert.deepEqual(data, { zeroWidth: [''] });
});
it('resolves a relative commons.root policy value against targetPath', () => {
mkdirSync(POLICY_DIR, { recursive: true });
writeFileSync(POLICY_FILE, JSON.stringify({ commons: { root: './fixture-commons' } }));
writeArtifact('lexicon/injection-lexicon', { version: '0.1.0', terms: [] });
const data = loadArtifact('lexicon/injection-lexicon', { targetPath: TEST_ROOT, fallback: null });
assert.deepEqual(data, { version: '0.1.0', terms: [] });
});
it('an explicit commonsRoot option takes precedence over the commons.root policy value', () => {
mkdirSync(POLICY_DIR, { recursive: true });
writeFileSync(POLICY_FILE, JSON.stringify({ commons: { root: join(TEST_ROOT, 'does-not-exist') } }));
writeArtifact('lexicon/injection-lexicon', { version: '0.1.0', terms: ['override-wins'] });
const data = loadArtifact('lexicon/injection-lexicon', {
targetPath: TEST_ROOT,
commonsRoot: FIXTURE_COMMONS_ROOT,
fallback: null,
});
assert.deepEqual(data, { version: '0.1.0', terms: ['override-wins'] });
}); });
it('defaults the fallback to null when the caller passes none', () => { it('defaults the fallback to null when the caller passes none', () => {
const data = loadArtifact('lexicon/injection-lexicon', { commonsRoot: FIXTURE_COMMONS_ROOT }); const data = loadArtifact('lexicon/injection-lexicon', { commonsRoot: FIXTURE_COMMONS_ROOT });
assert.equal(data, null); assert.equal(data, null);
}); });
it('freezes a loaded artifact so a caller mutation throws instead of leaking process-wide', () => {
writeArtifact('lexicon/injection-lexicon', { version: '0.1.0', terms: ['a'] });
const data = loadArtifact('lexicon/injection-lexicon', { commonsRoot: FIXTURE_COMMONS_ROOT });
assert.throws(() => { data.terms.push('mutated'); }, TypeError);
assert.throws(() => { data.version = '9.9.9'; }, TypeError);
});
it('does not read .llm-security/policy.json from any target — commons.root is not honored', () => {
// Simulates a hostile scanned target shipping .llm-security/policy.json
// with {"commons":{"root": "..."}} to redirect detection data. Even
// with such a file on disk and CLAUDE_PROJECT_ROOT pointed at it, the
// loader must ignore it entirely and fall through to the caller's
// explicit commonsRoot / default, never to policy.
const hostileTargetRoot = join(TEST_ROOT, 'hostile-target');
const hostilePolicyDir = join(hostileTargetRoot, '.llm-security');
mkdirSync(hostilePolicyDir, { recursive: true });
writeFileSync(
join(hostilePolicyDir, 'policy.json'),
JSON.stringify({ commons: { root: join(TEST_ROOT, 'attacker-controlled-empty-commons') } }),
);
const prevProjectRoot = process.env.CLAUDE_PROJECT_ROOT;
process.env.CLAUDE_PROJECT_ROOT = hostileTargetRoot;
try {
writeArtifact('lexicon/injection-lexicon', { version: '0.1.0', terms: ['still-here'] });
const data = loadArtifact('lexicon/injection-lexicon', { commonsRoot: FIXTURE_COMMONS_ROOT, fallback: null });
assert.deepEqual(data, { version: '0.1.0', terms: ['still-here'] });
} finally {
if (prevProjectRoot === undefined) delete process.env.CLAUDE_PROJECT_ROOT;
else process.env.CLAUDE_PROJECT_ROOT = prevProjectRoot;
}
});
}); });