Capability discovery
MCP, network, process execution, filesystem writes, credentials and approval controls.
ARL Guardian · Public Preview
See what authority surfaces exist in an AI-agent repository.
Give Guardian a repository. It scans the source without running the agent and maps evidence for MCP, outbound network access, filesystem mutation, process execution, credential references and approval boundaries.
Runs locally. No source upload. No LLM. No API key. No Guardian account.
RC5
Install and run it locally.
Requirement: Node.js 22.5 or later.
Install:
npm install https://agentrisklayer.com/downloads/agentrisklayer-guardian-0.3.0-rc.5.tgzGenerate a customer-facing report:
npx --no-install arl-guardian report /path/to/your-agentGenerate the technical evidence appendix:
npx --no-install arl-guardian evidence /path/to/your-agentSHA-256: eae7836e88e30902b13774f1e2b86d116d371d4c2ae3c30dd4fba71b7cf0ca35
What it gives you
A capability map with evidence, not a fake security score.
Guardian reduces a repository into a review surface so developers and assessors can see where potentially important authority lives.
Context separation
Separates agent-linked evidence from backend code, examples, tests, documentation and tooling.
Exact evidence
Keeps file and line references so a human can verify the source directly.
ARL handoff
Produces review questions and conservative candidate facts for human confirmation.
Real-world validation
Tested against a real public agent repository.
Browser Use · RC3 validation
Scanned without truncation.
Across six capability families.
110 refs, 13 agent-linked for human review.
24 refs, 0 agent-linked after removing a false positive.
Need the result reviewed?
Turn Guardian evidence into an ARL assessment.
Guardian observations are not findings, severity decisions or deployment verdicts. AgentRiskLayer reviews applicability, validates evidence, performs authorised testing where required, and retests affected controls.