Critical-review §4 E17 finding: pre-v7.2.0 the delegation-after-input
advisory fired only within a 5-call window. Attackers who deliberately
waited 6+ calls before delegating bypassed detection. Window was also
hardcoded — operators couldn't tune it for their environment.
Two coordinated changes:
1. LLM_SECURITY_ESCALATION_WINDOW env var (primary window override)
- parseInt(env) || getPolicyValue('trifecta', 'escalation_window', 5)
- Mirrors the established pattern from
LLM_SECURITY_TRIFECTA_MODE et al.
- Setting env=3 narrows; env=8 expands.
2. Secondary 20-call MEDIUM advisory (slow-burn variant)
- DELEGATION_ESCALATION_WINDOW_MEDIUM = 20 (hardcoded — same value
for all operators; tunable in a future patch if needed)
- checkEscalationAfterInput now returns `tier: 'primary'|'secondary'|null`
- formatEscalationWarning emits a different message for secondary —
mentions "slow-burn", references env-var, distinct from the
primary "DeepMind Category 4" framing
Hook reads max(WINDOW_SIZE, secondary+5) entries to cover the wider
window. Existing duplicate-suppression (`escalation_warning` state
entry) covers both tiers. Audit-trail event captures `tier` field.
Tests: +5 cases in tests/hooks/post-session-guard.test.mjs:
- secondary window catches 9-call distance (slow-burn)
- secondary boundary at exactly 20 calls
- primary regression guard (1-call distance)
- env=3 narrows primary (4-call distance becomes secondary)
- env=8 expands primary (7-call distance stays primary)
Updated existing test "does NOT trigger when input_source is >5 calls
ago" — now requires >20 calls (secondary window catches 6-20).
Suite: 1644 → 1672 (+28 from new tests + extended scope). All green.
CLAUDE.md hooks table updated to document both windows and the env var.
Closes A2 of v7.1.0 critical-review patch (docs/critical-review-2026-04-20.md):
- B4 (severity JSDoc): 4 critical = 93, not 90. Fixed in scanners/lib/severity.mjs:23
and CHANGELOG.md v7.0.0 tier description. The actual computation has always been
93 (70 + log2(5)*10 = 93.22 → round); only the docs were wrong.
- §5.4 co-monotonicity: new sweep test in tests/lib/severity.test.mjs over 15
representative count vectors. Asserts that (verdict, riskBand) agree under the
v7.0.0 contract for every case — catches future drift between riskScore tiers,
verdict cutoffs, and riskBand cutoffs. Includes a B4 anchor test (riskScore
{critical: 4} === 93) so doc/code drift fails loudly.
- B8 (CaMeL claims toned down): post-session-guard.mjs:646 comment block and
CLAUDE.md:184 Defense Philosophy bullet now describe the implementation
honestly — opportunistic byte-matching of truncated output fingerprints
(first 200 bytes, SHA-256/16-hex), not semantic data-flow tracking.
Trivially bypassed by mutation, summarisation, or re-encoding. Inspired by
CaMeL (DeepMind 2025), but not a CaMeL capability-tracking implementation.
Tests: 1495 → 1511 (+16: 15 sweep cases + 1 B4 anchor). All green.
Previously, `LLM_SECURITY_TRIFECTA_MODE=block` only exited 2 when the
detected trifecta was MCP-concentrated (all three legs via the same MCP
server) or involved sensitive-path + exfil. Distributed trifectas —
three legs originating from different tools, with a non-sensitive data
path and a non-sensitive exfiltration sink — were detected and warned
but not blocked. This mismatched the documented semantics of block mode
and gave operators a false sense of enforcement.
Change: remove the `(mcpInfo.concentrated || sensitiveExfil)` AND-gate
in the `TRIFECTA_MODE === 'block'` branch so any detected trifecta
blocks in block mode. Audit event `severity` still differentiates
critical (concentrated / sensitive-exfil) from high (distributed); the
blocked stderr message now explicitly names "Distributed trifecta:
three legs from different sources" when the confidence sub-signals
are absent.
Addresses critical review 2026-04-20 §2 B2 (HIGH) and §9 row 1
("enforces the Rule of Two").
Tests: 1 added (distributed trifecta in block mode now exits 2).
All 1495 tests pass.
New policy-loader.mjs reads .llm-security/policy.json with deep-merge against
defaults that exactly match existing hardcoded values. Integrated into all 7 hooks:
- pre-prompt-inject-scan: injection.mode (env var still takes precedence)
- post-session-guard: trifecta.mode, window_size, long_horizon_window
- pre-edit-secrets: secrets.additional_patterns
- pre-bash-destructive: destructive.additional_blocked
- pre-write-pathguard: pathguard.additional_protected
- pre-install-supply-chain: supply_chain.additional_blocked_packages
- post-mcp-verify: mcp.volume_threshold_bytes, mcp.trusted_servers
Backward compatible: no policy file = identical behavior to v5.1.0.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New audit-trail.mjs writes structured events to LLM_SECURITY_AUDIT_LOG path.
Integrated into post-session-guard at 6 warning emission points: trifecta,
escalation-after-input, data flow, volume threshold, slow-burn, behavioral drift.
No-op when env var not set — zero overhead for existing users.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
