DeepSafe Scan — Agent Instructions

DeepSafe Scan — Agent Instructions

This is the **deepsafe-scan** security scanner for AI agent environments. Works with: OpenClaw, Claude Code, Cursor, Codex, and any AI agent.

What this tool does

Scans for security risks in:

• AI agent config files (.claude/settings.json, .cursorrules, .vscode/tasks.json, etc.)
• Installed skills and MCP servers
• Session/memory data for leaked secrets and PII
• Deployment configuration
• Model behavioral safety (optional, needs API key)

When to use

Use this scanner when:

• The user asks about security, audit, or vulnerability scan of their AI setup
• The user is about to install a new skill or MCP server
• The user clones a new project with agent config files
• The user reports unexpected behavior (possible prompt injection)

How to invoke

**Static scan (no API key needed):**

python3 scripts/scan.py --modules posture,skill,memory,hooks --scan-dir . --no-llm --format markdown

**Full scan (auto-detects ANTHROPIC_API_KEY or OPENAI_API_KEY):**

python3 scripts/scan.py --modules posture,skill,memory,hooks,model --scan-dir . --format markdown

**OpenClaw users:**

python3 scripts/scan.py --openclaw-root ~/.openclaw --format html --output /tmp/report.html

**Targeted hooks-only scan (fastest):**

python3 scripts/scan.py --modules hooks --scan-dir . --no-llm --format markdown

Modules

Module	Scans	API Required
posture	openclaw.json or .env config files	No
skill	Skills, MCP servers — 15+ static analyzers	No (LLM analysis optional)
memory	Session data, credentials — 27+ secret patterns, 9 PII patterns	No
hooks	AI agent config files — 12 injection patterns	No
model	Behavioral safety probes (persuasion, sandbagging, deception, hallucination)	Yes

Interpreting results

**Scores:** Each module scores 1-100 (100 = clean). Total score is the average across active modules.

Score	Risk
85–100	LOW
65–84	MEDIUM
40–64	HIGH
1–39	CRITICAL

**After scanning:** Present CRITICAL and HIGH findings first. Explain the risk and remediation for each. Offer to fix issues.

LLM auto-detection

The scanner auto-detects API credentials in this order:

1. --api-base / --api-key flags
2. OpenClaw Gateway (~/.openclaw/openclaw.json)
3. ANTHROPIC_API_KEY environment variable
4. OPENAI_API_KEY environment variable
5. No API → static analysis only (all modules except model probes still work)