Verified Skill Auditor
Audit third-party agent skills for prompt injection, credential theft, and hidden backdoors before installation.
Free
Works with the AI Agents you already use
verified-skill-auditor
Example session with this skill installed
python3 scripts/audit.py scan samples/risky-skill/ --fail-on high
- Read your context and instructions
- Compiled the verified--auditor
The scan of samples/risky-skill/ identified 7 critical and 3 high severity findings. The audit has failed.
Audit Report: samples/risky-skill/
| Finding | Severity | File | Line | Explanation |
|---|---|---|---|---|
exec-pipe | CRITICAL | SKILL.md | 11 | Download piped directly into shell detected. |
injection | CRITICAL | SKILL.md | 15 | Instruction-override phrase "Ignore all previous instructions" found. |
hidden-instruction | CRITICAL | SKILL.md | 17 | Instructions hidden inside HTML comments. |
cred-access | CRITICAL | SKILL.md | 21 | Attempt to read SSH private keys (id_rsa). |
exfil | CRITICAL | SKILL.md | 23 | Local .env file being sent to external host. |
persistence-memory | HIGH | SKILL.md | 24 | Write operation to MEMORY.md detected. |
persistence-startup | MEDIUM | SKILL.md | 25 | Write to .bashrc for persistence. |
obfuscated | HIGH | SKILL.md | 26 | Base64 encoded payload piped to shell. |
net-unknown | HIGH | SKILL.md | 29 | Fetch from unknown/untrusted IP address 192.0.2.50. |
secret-leak | HIGH | SKILL.md | 31 | Hardcoded service key or token found in text. |
hidden-text | HIGH | SKILL.md | 34 | Invisible characters hiding text from reviewer. |
Verdict: FAIL (Exit Code 2)
Next steps
- Do not install this skill; the findings indicate high risk of credential theft and command injection.
- Remove the malicious
.bashrcandMEMORY.mdentries if you have already interacted with this code. - Review
TRIAGE.mdto understand thelethal-trifectaof injection, credential access, and network usage.
Connects securely to your tools. The creator never sees your data.
About this skill
Agentic skills run with access to your terminal and files, but most registries do not scan what they list. This auditor provides a local triage layer to detect prompt injection, credential theft, and hidden backdoors before you install third-party code. It scans skill bundles for patterns found in the ClawHavoc and ToxicSkills campaigns, mapping findings to the OWASP Agentic Skills Top 10.
What it does
- Scans for injections by identifying instruction-override phrases and hidden HTML comments meant to hijack the agent.
- Detects credential theft by flagging unauthorized reads of SSH keys, cloud credentials, and .env files.
- Blocks exfiltration by identifying local files being sent to unknown hosts via uploads or DNS tricks.
- Finds backdoors by catching writes to shell startup files, scheduled tasks, or memory files that survive restarts.
- Uncovers obfuscation by decoding base64, hex, and character-code blobs intended to hide malicious payloads.
- Flags broad access by reporting claims of unrestricted filesystem or shell permissions.
How it works
- Target the skill by providing the local path to the skill directory or a specific file.
- Execute the scan using the
scancommand to trigger the 14-rule pattern matcher. - Review the verdict by checking the pass, warn, or fail status alongside file and line-level evidence.
- Export results to JSON for CI/CD integration or use the
--fail-onflag to set severity thresholds.
Frameworks & tools
Built with Python for local execution. The rule engine targets patterns across Shell scripts, Markdown, and configuration files without requiring network access or API keys.
Why this beats prompting it yourself
Manual review is prone to missing invisible characters, encoded blobs, and hidden HTML comments that a human eye might skip. This skill automates the detection of known attack patterns from the 2026 threat landscape in seconds.
Use cases
- Auditing a new skill downloaded from a public repository before installation.
- Integrating a security gate into a skill development CI/CD pipeline.
- Triaging large skill bundles to prioritize manual review of high-risk files.
- Verifying that a local skill has not been tampered with by checking for persistence triggers.
Known limitations
Does not catch paraphrased natural language manipulation or novel obfuscation methods like ROT13. It lacks data-flow analysis and cannot detect payloads that only execute at runtime.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
No data-flow analysis or runtime behavioral checks. Limited to English instructions. Skips binaries and archives. Cannot verify publisher identity or cross-skill composition risks.
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 buying a skill.
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
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 6 days ago
- Free with an Agensi account
- Accounts
GitHub
Sil
Gnu
Apache
Example
- Software
Python
Permissions
Allowed Hosts