Open-source security gate for AI skills

Audit every skill before your agent runs it.

An open-weight gate that scans the skills you give Claude Code and Codex for prompt injection, then lets only the safe ones through.

Untrusted skills
pdf-extract
git-helper
meeting-sync
Jenz auditscan · classify ·
score on host
Your agent
pdf-extract
git-helper
meeting-sync

The skills nobody audits

36.8%of audited skills carried security flaws1,467 of 3,984
13.4%shipped a critical-severity issue534 skills
76were confirmed-malicious payloads8 still live at disclosure

91% of the malicious skills worked through prompt injection. Source: Snyk ToxicSkills audit of 3,984 published skills, Feb 2026.

How it works

One pipeline. The model advises, your host decides.

Four stages turn an untrusted file into a verdict you can trust.

Step 01

Scan

A regex and Unicode pass flags the obvious stuff: exfil URLs, shell payloads, hidden characters.

Step 02

Classify

An open-weight model reads it as inert data and returns its findings twice, for consistency.

Step 03

Decide

A function on your host turns findings into the verdict. The model gets no vote.

Step 04

Gate

Safe skills load. Everything else is blocked with a reason. Errors fail closed.

Inside the audit

For every skill, Jenz spawns a panel of security agents.

Each agent reads the skill as inert data, quoted and never executed, and hunts a different class of attack against a known threat catalog. They return findings, not verdicts.

SKILL.mduntrusted
Data & secretsAgent 01

Hunts attempts to read credentials, env files, or your repo and ship them off-box.

OWASP LLM06ATLAS · Exfiltration
Commands & toolsAgent 02

Flags shell payloads, destructive actions, and tool or permission abuse.

OWASP AgenticATLAS · Execution
Instructions & controlAgent 03

Catches prompt injection, jailbreaks, and “ignore previous instructions” overrides.

OWASP LLM01
Security model

Built to be trusted, not just clever.

Every finding maps to the attack frameworks the industry already trusts, OWASP and MITRE. The verdict is computed on your host, never by the model. Safe is something a skill earns.

Don’t trust a skill you haven’t audited.

Open-source, open-weight, fail-closed. Run your first skill through the gate in under a minute.

Built atAda Ventures Open-Source AI HackathonSafety, Security & Governance