agent-supply-chain-guard

Agent Supply Chain Guard

Before your agent reads it, scan it.

Your AI agent can be compromised by a file it reads.

An innocent-looking SKILL.md, MCP manifest, plugin, or README can contain instructions that redirect an agent, expose credentials, run commands, or request far more access than it needs.

Agent Supply Chain Guard finds those signals before they reach your agent runtime. It is free, local-first, dependency-free, and takes one command to run.

See it catch a poisoned skill in 10 seconds

git clone https://github.com/ppradyoth/agent-supply-chain-guard.git
cd agent-supply-chain-guard
python -m agent_supply_chain_guard scan examples/poisoned-skill.md

Expected output:

examples/poisoned-skill.md:5: [hidden-instruction] instruction-like text found in supply-chain content
examples/poisoned-skill.md:6: [process-execution] process execution reference
examples/poisoned-skill.md:7: [unrestricted-permission] broad or unrestricted permission
3 finding(s)

Install and scan in 30 seconds

pipx install agent-supply-chain-guard
agent-supply-chain-guard scan .

Or run it without installing:

python -m pip install agent-supply-chain-guard
agent-supply-chain-guard scan path/to/agent-project

A clean scan exits 0. A scan with security signals exits 1, so it works naturally in CI:

agent-supply-chain-guard scan . --quiet

GitHub Action

name: Agent security
on: [push, pull_request]
permissions: {}
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: ppradyoth/agent-supply-chain-guard@v0.1.0
        with:
          path: .

Pin third-party actions to reviewed commits for production use.

What it catches

JSON output is available for automation:

agent-supply-chain-guard scan . --format json

Optional GPT or Claude review

Rules are the private, free baseline. If you want a second-pass explanation and prioritization, install one provider extra and export your own key in the shell:

pipx install 'agent-supply-chain-guard[openai]'
export OPENAI_API_KEY='your-key'
agent-supply-chain-guard scan . --ai-provider openai

Claude is the equivalent:

pipx install 'agent-supply-chain-guard[anthropic]'
export ANTHROPIC_API_KEY='your-key'
agent-supply-chain-guard scan . --ai-provider anthropic

Use AGENT_GUARD_OPENAI_MODEL or AGENT_GUARD_ANTHROPIC_MODEL to choose a model, or pass --model. The key is read from the environment and never printed or saved. Only matched finding evidence is sent for review; do not use AI mode on sensitive repositories unless that disclosure is acceptable. Provider fees and retention policies apply. AI output is advisory and does not alter the deterministic exit status.

Important limitation

This is a signal scanner, not a proof of exploitability. Findings need human review, and a clean result does not make an agent or dependency safe. The goal is to make the first security check cheap enough that teams actually run it.

Why now?

Agent security has moved beyond chatbot jailbreaks. MCP servers, agent skills, plugins, OAuth grants, and tool calls are becoming a software supply chain—and every new connection adds another place for untrusted instructions or excessive authority to enter.

Contributing

Add a narrowly scoped rule with a harmless fixture and a test. See CONTRIBUTING.md. Please never commit real credentials or client data.

Licensed under Apache-2.0.