ai-safe2-framework

Combined Threat Model: AI SAFE² × SlowMist for OpenClaw

This document defines the combined threat model for an OpenClaw deployment secured by both the SlowMist Security Practice Guide v2.7 and the AI SAFE² Framework. It extends SlowMist’s defined scope with AI SAFE²’s organizational and cross-deployment threat surface.


Deployment Assumptions

Both frameworks target deployments where:


Threat Categories

T1 — Prompt Injection (Cognitive Layer)

Description: Malicious instructions embedded in documents, emails, web content, or tool outputs that cause the agent to execute unintended actions.

SlowMist coverage: Red-line rules block the most dangerous outcomes; yellow-line confirmation gates catch high-risk results of successful injection.

AI SAFE² coverage: Memory Vaccine encodes injection-detection heuristics as persistent cognitive context; Control Gateway intercepts injection payloads at the API layer before they reach the model.

Residual risk: Highly sophisticated injection payloads that evade both the cognitive rules and the gateway’s pattern matching. Mitigation: quarterly red-team exercises specifically targeting injection resistance.


T2 — Supply Chain Poisoning (Skill/MCP Layer)

Description: A malicious or tampered Skill, MCP, or script is installed, introducing a backdoor, exfiltration channel, or destructive capability into the agent’s trusted computing base.

SlowMist coverage: Offline audit protocol requires full-text scan of every skill before installation; human approval gate required.

AI SAFE² coverage: Vulnerability Scanner runs recurring post-installation checks to detect skills added outside the approval workflow; Signed Skills (roadmap) will provide cryptographic verification.

Residual risk: Highly obfuscated malicious code (cf. CVE-2024-3094 / xz-utils) that survives full-text scanning. Mitigation: SlowMist’s human review step is the last line of defense for sophisticated obfuscation.


T3 — Persistent Memory Poisoning (Vector DB / Long-Term Memory Layer)

Description: An attacker (via prompt injection, malicious skill output, or direct access) writes malicious “facts” into OpenClaw’s long-term memory, causing persistent behavioral modification across sessions.

SlowMist coverage: Not explicitly addressed in the v2.7 guide.

AI SAFE² coverage: Memory Vaccine is specifically designed for this threat — it inspects and filters inputs to long-term memory for malicious or misaligned content and encodes anti-poisoning heuristics as higher-priority persistent context.

Residual risk: Memory poisoning payloads designed to appear benign to the Memory Vaccine’s heuristics. Mitigation: human review of trusted-source and persistent-instruction lists (weekly); Memory Vaccine updates on discovery of new poisoning patterns.


T4 — Destructive Operations (Execution Layer)

Description: The agent executes irreversible destructive actions (filesystem destruction, credential exfiltration, data wipe) either through successful attack or agent reasoning error.

SlowMist coverage: Red-line rules prevent the most dangerous individual commands; yellow-line gates require human confirmation for high-risk operations; brain backup enables state recovery.

AI SAFE² coverage: Control Gateway enforces high-risk tool denial externally, independent of agent self-compliance; circuit-breaker configuration provides automated halts before human review.

Residual risk: Same-UID execution (SlowMist-acknowledged limitation) — malicious code running in the agent’s UID can still act on files accessible to that user. Mitigation: dedicated OS user per agent + container isolation (increases complexity but closes this gap).


T5 — Credential Exposure (Secrets Layer)

Description: API keys, OAuth tokens, private keys, or mnemonic phrases stored in OpenClaw’s memory, logs, or config files are exposed via direct access, exfiltration, or backup compromise.

SlowMist coverage: Nightly audit metrics 10 and 11 scan for plaintext credentials and environment variable exposure; credential/state separation guidance discourages storing keys in behavioral backups.

AI SAFE² coverage: Vulnerability Scanner’s secret hunter actively scans all files (logs, history, config) for high-entropy strings matching key patterns; Memory Vaccine blocks redaction bypass.

Residual risk: Novel credential formats not in Scanner pattern library; credentials injected into memory after last nightly scan (24h gap). Mitigation: run Scanner after any event that could introduce new files.


T6 — Unauthorized Network Access / SSRF (Network Layer)

Description: The agent is induced to make unauthorized network requests — exfiltrating data, reaching internal infrastructure, or serving as a pivot point.

SlowMist coverage: Nightly process/network audit checks for anomalous outbound connections and unexpected listeners; dedicated VM isolation recommended.

AI SAFE² coverage: Control Gateway enforces approved egress domain allowlist externally; Vulnerability Scanner checks for admin panel binding to 0.0.0.0; OpenClaw v2026.2.13 patched one SSRF vector (link extractor internal IP blocking).

Residual risk: SSRF vectors in newly installed skills or OpenClaw updates not yet patched. Mitigation: Gateway external allowlist is the durable control — blocks unapproved egress regardless of what code runs inside the agent.


T7 — Cross-Agent Impersonation (Multi-Agent Layer)

Description: In multi-agent or A2A orchestration scenarios, a malicious agent impersonates a trusted peer to inject instructions into OpenClaw’s reasoning chain.

SlowMist coverage: Not addressed in v2.7 (single-agent scope).

AI SAFE² coverage: Engage & Monitor pillar identifies anomalous patterns in agent-to-agent communication; semi-annual A2A impersonation red-team exercise codified in red-team-schedule.md.

Residual risk: Novel A2A attack patterns not yet in threat model. Mitigation: annual threat model review incorporating emerging multi-agent attack research.


T8 — Organizational / Fleet Compromise (Ecosystem Layer)

Description: An attacker who compromises one OpenClaw instance uses it as a pivot to reach others across the organization (credential reuse, shared network access, shared backup repositories).

SlowMist coverage: Not addressed (single-host scope).

AI SAFE² coverage: Cross-deployment inventory enables identification of shared credentials and blast-radius assessment; fleet-wide anomaly detection surfaces lateral movement signals; deployment registry documents trust isolation between instances.

Residual risk: Insufficient isolation between deployments (shared credentials, shared backup repos). Mitigation: enforce credential uniqueness per deployment; separate backup repos per instance; network-segment agent hosts.


Explicit Out-of-Scope Items

The following are acknowledged limitations of the combined framework. They represent known residual risks that require additional controls or architectural changes outside this framework’s current scope:


This threat model should be reviewed annually and updated to incorporate new OpenClaw releases, new SlowMist guide versions, and emerging attack research. See red-team-schedule.md for the exercise cadence built around this threat model.