Scan AI skills.
Before your agent runs them.
A useful skill can also bring scripts, downloads and access to sensitive files. Pwnry inspects your skill ZIP before you install it, with a private evidence report instead of a blind leap of trust.
For teams using Claude Code, Codex, Gemini and other agents with file-based skills. Fully automated analysis, without manual review.
ZIP or public GitHub skill. Up to 16 MiB. Private by default. Your allowance appears after sign-in.
A download-and-execute pattern needs an explanation. The report names the file and the evidence. An indicator alone does not prove malicious intent.
What is a Skill Check?
A Skill Check is an automated security inspection of an AI agent skill package. It combines a file inventory, code indicators and bounded analysis of selected executable content. The result documents findings, observed behavior and what could not be checked.
Understand Claude Code and Codex skill risksFrom ZIP to a decision.
The report stays tied to the exact package hash. A changed version needs a new inspection.
-
Submit the complete package
Upload a ZIP or import a public GitHub skill folder. SKILL.md, scripts and bundled resources are included before analysis. Dependencies are not installed. Only submit packages you are allowed to analyse, without real credentials.
-
Inspect code and selected behavior
Rules look for suspicious references and execution patterns. Selected scripts and binaries receive deeper isolated analysis without network access. Missing dependencies and unsupported runtimes remain visible.
-
Read the evidence before installing
Your private report links findings to files and observed activity. Fix or clarify suspicious behavior, then inspect the changed package again. Share a separate report only when you choose to.
Useful evidence. Clear limits.
A reference to a credential file is a reason to investigate, not proof of data theft. A blocked connection is an observed attempt, not a successful transfer. Pwnry keeps those distinctions in the report.
The current check does not call a language model and does not recreate a complete Claude Code, Codex or Gemini session. It does not fetch remote payloads. No findings therefore means no findings within the documented coverage, not a malware-free certificate.
Research behind PwnryWhat a Skill Check will cost.
Coming soonPlanned base price: 1 € net per check. One submission costs one unit, including the deeper analysis. A technical failure returns the unit.
| Package | Skill Checks | Of which bonus |
|---|---|---|
| €10 | 11 | +1 |
| €25 | 28 | +3 |
| €50 | 60 | +10 |
| €100 | 125 | +25 |
All prices net, plus statutory VAT.
Good questions.
Clear answers.
What you should know before your first check.
What does the Pwnry Skill Scanner check?+
Pwnry inventories the submitted ZIP, hashes its files and checks text for indicators such as credential references, instruction overrides, encoded execution and download-and-execute patterns. Selected scripts and binaries receive bounded isolated analysis. Findings and coverage limits appear in a private report.
Can I check Claude Code, Codex and Gemini skills?+
You can submit their skill files as a ZIP, including SKILL.md and bundled scripts or binaries. Pwnry examines the package, not a complete running session of the target agent. Unsupported runtimes and missing dependencies are reported as limitations.
Does the analysis send my skill to an AI model?+
The current Skill Check uses deterministic rules, binary analysis and isolated execution. It does not call a language model. This is separate from AI-assisted penetration testing elsewhere on Pwnry.
Who can see my uploaded skill and report?+
Results are private to your account by default. Sharing creates a separate report without uploaded files, source paths or raw tool output. An unlisted report can be read by anyone with its link; a published report may be indexed. Only share material you are allowed to disclose.
Is a skill with no findings guaranteed safe?+
No. A result applies to the submitted package hash and the analysis that actually ran. Network access is disabled; remote payloads and dependencies are not downloaded. Context-dependent prompt injection and the behavior of a complete agent session are not fully tested. Read the documented limitations before installation.
How do I start and what does a check cost?+
Sign in to see your current beta allowance and available checks. Promotional vouchers can add Skill Check units. One submitted ZIP uses one unit, including its deeper analysis and worker retries; technical failures restore the unit. Paid package checkout is not available yet.
Check the package before trusting it.
Start with your current beta allowance or redeem an event voucher. The available checks appear in your account.
Open Skill Checks