llm-security/scanners/lib/policy-loader.mjs
Kjell Tore Guttormsen f3aaf5479f fix(llm-security): scanner false-positives/negatives — trigger, toxic-flow, policy-loader (#38-#41,#57,#26)
#38 toxic-flow matched trifecta-leg keywords with bare includes(), so substrings ('url' in 'curl', 'key' in 'monkey', 'auth' in 'author') fabricated CRITICAL trifectas on benign components; now word-boundary matched. #40 TRG-broad fired HIGH on a bare any/all/every anywhere ('fix any lint errors'); the universal-claim regex now requires genuine universal phrasing. #41 TRG-baiting substring-matched ('any file' in 'many files'); now boundary-anchored. #39 the broad-name list missed multi-char generic names (helper/assistant/auto/general/agent/tool); widened coherently so it does not reintroduce #40. #57 the '(recovered from obfuscation)' label compared raw against a lowercased normal form, firing on any uppercase char; now gated on an explicit decode-changed flag.

#26 (same file) getPolicyValue used 'key in sectionObj' with no type guard, so a scalar section override in policy.json (e.g. {"injection":"block"}) threw an uncaught TypeError; now guarded to fall back to the default. Suite 1931/0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
2026-07-18 10:15:10 +02:00

256 lines
8.4 KiB
JavaScript

// policy-loader.mjs — Central policy file reader for distributable hook configuration
// Reads .llm-security/policy.json from project root. Falls back to defaults
// matching existing hardcoded behavior when no policy file exists.
// Zero external dependencies.
import { readFileSync } from 'node:fs';
import { join } from 'node:path';
// ---------------------------------------------------------------------------
// Default policy — matches all existing hardcoded values exactly
// ---------------------------------------------------------------------------
const DEFAULT_POLICY = Object.freeze({
version: '1.0',
injection: {
mode: 'block',
medium_advisory: true,
custom_patterns: [],
},
trifecta: {
mode: 'warn',
window_size: 20,
long_horizon_window: 100,
escalation_window: 5,
},
secrets: {
additional_patterns: [],
allowed_paths: [],
},
destructive: {
additional_blocked: [],
allowed_commands: [],
},
pathguard: {
additional_protected: [],
allowed_paths: [],
},
supply_chain: {
additional_blocked_packages: [],
trusted_registries: [],
},
mcp: {
trusted_servers: [],
volume_threshold_bytes: 100_000,
cumulative_drift_threshold: 0.25,
},
audit: {
log_path: null,
events: ['trifecta', 'injection', 'secrets', 'destructive'],
},
ci: {
failOn: null,
compact: false,
},
entropy: {
thresholds: {
critical: { entropy: 5.4, minLen: 128 },
high: { entropy: 5.1, minLen: 64 },
medium: { entropy: 4.7, minLen: 40 },
},
// User-extensible extension skip list — merged with built-in defaults.
suppress_extensions: [],
// Additional line-level regex sources (string or array of strings compiled at load).
suppress_line_patterns: [],
// Substring matches against relative path — plain contains, no glob.
suppress_paths: [],
},
// TRG — trigger/activation-abuse scanner. Lists mirror the scanner defaults
// in scanners/trigger-scanner.mjs; override any of them via policy.json.
trg: {
mode: 'warn',
baiting_phrases: [
'anything', 'everything', 'always', 'whenever', 'no matter what',
'any request', 'any task', 'any file', 'every time',
'all files', 'all messages', 'all requests',
],
builtin_names: [
'read', 'write', 'edit', 'bash', 'glob', 'grep', 'task',
'webfetch', 'websearch', 'notebookedit', 'todowrite',
'ls', 'cat', 'agent', 'search', 'fetch',
],
broad_single_words: [
'run', 'do', 'go', 'help', 'fix', 'use', 'get', 'set', 'all', 'any', 'it', 'this', 'that',
'helper', 'assistant', 'auto', 'general', 'agent', 'tool',
],
},
// SIG — known-bad-identity signature engine. Toggle families or point at a
// custom ruleset via policy.json.
sig: {
enabled_families: ['webshell', 'reverse_shell', 'cryptominer', 'hacktool'],
custom_rules_path: null,
},
// AST — Python AST taint scanner (shells out to a parse-only python3 helper).
ast: {
enabled: true,
python_path: 'python3',
timeout_ms: 5000,
},
});
// Cache loaded policy per project root
const cache = new Map();
// Module-scoped Set of env-var names already warned about — dedupes to one
// stderr line per env-var per process, regardless of how many call-sites
// invoke getPolicyValueWithEnvWarn for the same name.
const _warnedEnvVars = new Set();
/**
* Resolve project root from env or cwd.
* @param {string} [explicitRoot]
* @returns {string}
*/
function resolveRoot(explicitRoot) {
return explicitRoot || process.env.CLAUDE_PROJECT_ROOT || process.cwd();
}
/**
* Deep merge two objects (source overrides target).
* @param {object} target
* @param {object} source
* @returns {object}
*/
function deepMerge(target, source) {
const result = { ...target };
for (const key of Object.keys(source)) {
if (
source[key] !== null &&
typeof source[key] === 'object' &&
!Array.isArray(source[key]) &&
typeof target[key] === 'object' &&
!Array.isArray(target[key])
) {
result[key] = deepMerge(target[key], source[key]);
} else {
result[key] = source[key];
}
}
return result;
}
/**
* Load policy from .llm-security/policy.json.
* Returns defaults if no policy file exists or if parsing fails.
* Cached per project root (per process).
*
* @param {string} [projectRoot] - Explicit root, or derived from env/cwd
* @returns {object} Merged policy with defaults
*/
export function loadPolicy(projectRoot) {
const root = resolveRoot(projectRoot);
if (cache.has(root)) return cache.get(root);
const policyPath = join(root, '.llm-security', 'policy.json');
let policy;
try {
const raw = readFileSync(policyPath, 'utf-8');
const parsed = JSON.parse(raw);
policy = deepMerge(DEFAULT_POLICY, parsed);
} catch {
// No policy file or invalid JSON — use defaults
policy = { ...DEFAULT_POLICY };
}
cache.set(root, policy);
return policy;
}
/**
* Get a specific policy value with fallback.
* Environment variables ALWAYS take precedence over policy file values.
*
* @param {string} section - Policy section (e.g. 'injection', 'trifecta')
* @param {string} key - Key within section (e.g. 'mode', 'window_size')
* @param {*} defaultValue - Fallback if neither policy nor default has the value
* @param {string} [projectRoot] - Explicit root
* @returns {*}
*/
export function getPolicyValue(section, key, defaultValue, projectRoot) {
const policy = loadPolicy(projectRoot);
const sectionObj = policy[section];
// v7.8.3 (#26): a scalar section override in policy.json (e.g.
// {"injection": "block"}) survives deepMerge — guard before `in` so it
// falls back to the default instead of throwing a TypeError.
if (sectionObj && typeof sectionObj === 'object' && key in sectionObj) return sectionObj[key];
return defaultValue;
}
/**
* Resolve a policy value with an overlapping env-var, emitting a one-time
* stderr deprecation warning when both the env-var AND the policy.json key
* are explicitly set.
*
* Resolution order (env-wins is unchanged from getPolicyValue contract):
* 1. If LLM_SECURITY_DEPRECATION_QUIET=1, suppress warning logic entirely
* and return env-value (if defined) else policy-value.
* 2. Otherwise, if env-var is set AND policy-value differs from
* defaultValue (heuristic: user wrote the key in policy.json),
* emit one stderr warning per envVarName per process.
* 3. Return env-value if defined, else policy-value.
*
* Why "differs from defaultValue" rather than parsing the raw policy file:
* loadPolicy() deep-merges DEFAULT_POLICY so `key in policy[section]` is
* always true. Comparing the resolved value to the caller's defaultValue
* is a reliable proxy for "user explicitly overrode this in policy.json"
* because callers pass defaults that match DEFAULT_POLICY.
*
* @param {string} section - Policy section (e.g. 'injection', 'trifecta')
* @param {string} key - Key within section (e.g. 'mode')
* @param {string} envVarName - Overlapping env-var (e.g. 'LLM_SECURITY_INJECTION_MODE')
* @param {*} defaultValue - Hardcoded default (must match DEFAULT_POLICY value)
* @param {string} [projectRoot] - Explicit root
* @returns {*}
*/
export function getPolicyValueWithEnvWarn(section, key, envVarName, defaultValue, projectRoot) {
const envValue = process.env[envVarName];
if (process.env.LLM_SECURITY_DEPRECATION_QUIET === '1') {
if (envValue !== undefined) return envValue;
return getPolicyValue(section, key, defaultValue, projectRoot);
}
const policyValue = getPolicyValue(section, key, defaultValue, projectRoot);
if (envValue !== undefined && policyValue !== defaultValue) {
if (!_warnedEnvVars.has(envVarName)) {
_warnedEnvVars.add(envVarName);
process.stderr.write(
`[llm-security] Deprecation: env-var ${envVarName} will be removed in v8.0.0; ` +
`policy.json key ${section}.${key} also set — env wins for now. ` +
`Suppress with LLM_SECURITY_DEPRECATION_QUIET=1.\n`
);
}
}
if (envValue !== undefined) return envValue;
return policyValue;
}
/**
* Get the full default policy (for documentation/example generation).
* @returns {object}
*/
export function getDefaultPolicy() {
return JSON.parse(JSON.stringify(DEFAULT_POLICY));
}
/**
* Reset cache and warning dedup state (for testing only).
*/
export function _resetCacheForTest() {
cache.clear();
_warnedEnvVars.clear();
}