| 1 |
Behavioral Red Lines (absolute prohibitions: rm -rf /, blind hidden instruction execution, direct credential exfiltration) |
Pre |
S&I |
MV |
Cognitive-layer enforcement via agent’s own reasoning |
MV makes rules persistent across sessions; external Gateway enforces same rules outside agent’s cognition layer |
| 2 |
Behavioral Yellow Lines (pause + human confirm: sudo, SSH key ops, financial transactions, bulk deletion) |
Pre |
S&I, F&R |
GW |
Human-in-the-loop gate for high-risk operations |
GW adds automated pre-emptive blocking; doesn’t require the agent to self-police |
| 3 |
Skill Installation Audit (offline clone → full-text scan incl. Markdown/JSON → red-flag check → human approval) |
Pre |
S&I |
VS |
Supply-chain intake protocol; blocks malicious code before agent executes it |
VS scans existing installed skills on a recurring basis; SlowMist audits at install time only |
| 4 |
Permission Narrowing (agent self-constrains file and process permissions to minimum needed) |
In |
S&I |
GW |
Least-privilege execution at OS layer |
GW enforces least-privilege at the API/tool-call layer independently of agent self-compliance |
| 5 |
Hash Baseline for Critical Configs (openclaw.json, sshd_config, authorized_keys) |
In / Post |
A&I |
VS |
Configuration integrity; detects unauthorized file drift |
VS extends detection to secrets exposure and network bindings that hash comparison misses |
| 6 |
Audit Logging with chattr +i (immutable log attributes) |
In |
A&I, E&M |
GW |
Tamper-evident local audit trail |
GW provides a separate immutable API-layer audit log outside the agent’s filesystem, unaffected by a host compromise |
| 7 |
Cross-Skill Pre-flight Business Risk Checks (agent evaluates risk before acting across skill boundaries) |
In |
E&M |
GW |
Runtime behavioral monitoring at agent cognition layer |
GW performs same check externally at API layer; does not rely on agent’s self-assessment |
| 8 |
Nightly 13-Metric Audit (platform scan, processes, directory changes, cron integrity, SSH failures, hash baseline, yellow-line counts, disk, env vars, credential scan, skill baseline, backup confirmation) |
Post |
A&I |
VS |
Comprehensive host-layer posture assessment |
VS runs as a complementary scan (secrets, network exposure, permissions); outputs a risk score for trending; route both to centralized log for multi-deployment view |
| 9 |
Push Notification on Audit Completion (Telegram / Discord / Signal) |
Post |
F&R, E&M |
— |
Confirms audit pipeline is functioning; surfaces silent failures |
Centralized log aggregation enables cross-deployment alert correlation (AI SAFE² E&M) |
| 10 |
Report Persistence at /tmp/openclaw/security-reports/ |
Post |
F&R |
— |
Local fallback when push delivery fails |
Back up to centralized store for fleet-level retention and forensic availability |
| 11 |
Brain Backup to Private Repo (OpenClaw state directory pushed nightly) |
Post |
F&R |
— |
Behavioral state recovery; enables rollback to known-good cognitive context |
Separate credential rotation and backup per AI SAFE² Fail-Safe pillar; JIT credential issuance limits recovery-time exposure |
| 12 |
Credential/State Separation (explicit guidance: don’t store keys in behavioral backup) |
Dep |
F&R, S&I |
— |
Limits blast radius during recovery |
AI SAFE² formalizes this as an identity architecture requirement with automated rotation bots |
| 13 |
Dedicated VM / Isolated Host Recommendation |
Dep |
S&I |
— |
Workload isolation; trust boundary separation |
AI SAFE² extends to container-level isolation policies, network segmentation, and multi-tenant controls |
| 14 |
Same-UID Compromise Acknowledgment (documented limitation) |
— |
S&I |
GW |
Honest threat model scoping |
GW operates outside the agent’s UID trust boundary; closes this specific gap |
| 15 |
Security Validation & Red Teaming Guide (end-to-end test: cognitive injection → host escalation → exfiltration → persistence → audit tampering → recovery) |
Ongoing |
E&E |
— |
Comprehensive red-team curriculum covering all three defense tiers |
AI SAFE² E&E pillar schedules recurring exercises (quarterly/annual) on this curriculum; SlowMist provides content, AI SAFE² provides cadence |
| 16 |
Agent-Native Deployment (agent reads, evaluates, and self-deploys the defense matrix) |
Dep |
E&E |
MV |
Reduces human configuration error; builds operator intuition through agent self-assessment |
AI SAFE² formalizes as organizational training: operators learn from agent-deployment process |
| 17 |
Threat Model Documentation (explicit scope: what the guide does and does not protect) |
Dep |
E&E |
— |
Calibrates operator expectations |
AI SAFE² E&E requires annual threat model review incorporating new OpenClaw releases and CVEs |