Agent Hardening

Lock down any LLM agent against prompt injection, data exfiltration, social engineering, and channel-based attacks.

Built from real pen-test findings, not theory. Works with OpenClaw, Claude Code, LangChain, and any agent that takes natural-language input and calls external tools.

Quick start

1. Copy agent-hardening/ into your skills directory
2. Run the attack surface checklist (references/attack-surface-checklist.md)
3. Audit MCP connections (references/mcp-hardening.md)
4. Apply the tiered behavioral rules to your agent's operating docs
5. Verify with the automated test runner or the manual quick test
6. Fix failures, re-test, document findings

Automated testing

# Test against any OpenAI-compatible endpoint
python agent-hardening/tools/run-security-tests.py \
    --endpoint https://api.openai.com/v1/chat/completions \
    --api-key sk-... \
    --model gpt-4 \
    --owner-name "Don" \
    --output findings.json

# Test against local Ollama
python agent-hardening/tools/run-security-tests.py \
    --endpoint http://localhost:11434/v1/chat/completions \
    --model llama3

# Test your hardened system prompt
python agent-hardening/tools/run-security-tests.py \
    --endpoint https://api.openai.com/v1/chat/completions \
    --api-key sk-... \
    --model gpt-4 \
    --system-prompt-file my-agent-prompt.txt \
    --output findings.json

Requires Python 3.10+ and requests (pip install requests).

Included files

File	Purpose
SKILL.md	Skill entrypoint and workflow
references/attack-surface-checklist.md	Identify what the agent can access
references/channel-hardening.md	Per-channel security configuration
references/mcp-hardening.md	MCP server permission auditing
references/behavioral-rules.md	4-tier defensive operating rules
references/quick-test.md	10 single-shot + 5 multi-turn security tests
references/findings-template.md	Structured findings documentation
tools/run-security-tests.py	Automated test runner (10 single-shot tests)

What it covers

See SKILL.md for the full workflow, principles, and framework compatibility notes.

License

MIT