Skip to main content
AgentSecv0.3.1
OWASP AST-10 + Web3 Annexv0.3.1 · MIT

Audit every
Agent Skill

One command scans every skill installed in your project — for vulnerabilities, supply-chain risk, and policy drift. Results in seconds, aligned with the OWASP Agentic Skills Top 10.

10
OWASP AST categories
4
Output formats
3
Agent platforms
0–100
Skill score range
The audit, live

Zero install. Zero config. Just run it.

agentsec — ~/.openclaw

Auto-detects every skill · Scans against OWASP AST-10 · Exits with code 1 on policy violation

How it works

Three steps.
Seconds each.

AgentSec is a single binary with zero runtime dependencies. It runs locally, never phones home, and produces deterministic audit artifacts you can diff, sign, and replay.

01

Discover

Crawls .skill/, skill.md files, and platform manifests across your project. No config, no flags — it finds every installed skill.

openclaw · claude · codex
02

Audit

Runs each skill through 10 OWASP categories: malicious prose injection, over-permissioned scopes, supply-chain drift, weak isolation, and more.

10 categories · 47 rules
03

Report

Text for humans, JSON for scripts, SARIF for IDEs, HTML for stakeholders. Gate CI with the exit code — zero special flags required.

4 formats · 4 policy presets
OWASP Agentic Skills Top 10

The 10 ways agent skills fail.

AgentSec scans for each of them, every run. Click a category to read the spec.

AST-01Critical

Malicious Skills

What it is

Skills that look legitimate but ship hidden payloads — credential stealers, backdoors, prompt injection buried in prose.

Why it matters

Skills run with the agent's full permissions. One bad install can leak keys, SSH, wallets, and shell access.

agentsec --verbose | grep AST-01OWASP AST-01
Outputs

One scan.
Four formats.

Humans read text. CI reads JSON. IDEs consume SARIF. Stakeholders open HTML. AgentSec emits any of them with --format.

audit-report.txt--format text
Loading audit-report.txt
Policies

Pick a preset. Or write your own.

Every audit runs against a policy. Four presets ship built-in, and you can define custom gates in .agentsecrc.

--policy default
Balanced
Blocks critical:yesBlocks high:Requires tests:
Blocks critical findings.
--policy strict
Enterprise
Blocks critical:yesBlocks high:yesRequires tests:yes
Blocks critical & high, enforces tests.
--policy permissive
Dev
Blocks critical:CVE onlyBlocks high:Requires tests:
Only blocks critical CVEs.
--policy owasp-top-10
Compliance
Blocks critical:yesBlocks high:yesRequires tests:
Direct mapping to OWASP AST-10.
.agentsecrc
{
"policy": "strict",
"platform": "openclaw",
"format": "sarif",
"output": "audit.sarif"
}
Install

Audit skills.
Trust agents.

Start with npx agentsec. No install, no config, no flags.

Get in touch

Questions about
skill security?

Whether you're building a skill, evaluating skills you've installed, or thinking about agent security broadly, say hello.