More screenshots
Works with the AI Agents you already use
Skill Security Scanner: Check Skills Before Install
Turn any third-party skill or plugin into a SAFE / REVIEW / DANGEROUS verdict with file:line evidence, before it runs on your machine.
Free
About this skill
Installing a skill or plugin gives it your shell, your files and your keys. Malicious skills are already out there: hundreds of them used a fake "Prerequisites" step that decodes a hidden command into a shell. This scanner flags that lure, plus SSH-key theft, prompt-reading hooks and hidden prompt injection, with file:line evidence, before you install. On Anthropic's own 20 skills it gave 0 false DANGEROUS verdicts.
Skill Security Audit ships skill_audit.py, a local static scanner. It reads every file of a skill folder, plugin, marketplace repo or .zip. It never runs, imports or extracts anything. The verdict per skill is DANGEROUS / REVIEW / SAFE, and the agent knows what to do with each.
What it catches:
- Download-and-run lures: a download piped into a shell, a decoded command piped into an interpreter, encoded PowerShell, password-protected downloads.
- Credential theft: reads of SSH keys, cloud and developer credentials, browser profiles, wallets, shell history and agent transcripts. A file that reads secrets and calls the network is flagged as an exfiltration chain.
- Agent hooks and config: hooks that run on every tool call or see every prompt, hooks that run a networked script, unrestricted
allowed-tools, permission bypass flags, MCP servers that are unpinned, remote, or given the whole disk. - Hidden instructions: override and "don't tell the user" phrasing, hidden HTML comments, zero-width, bidi and Unicode tag characters (the hidden text is decoded and shown).
- Persistence and damage: shell startup files, cron and scheduled tasks, authorized_keys, disabling quarantine or antivirus, deleting the home folder.
- Opaque content: binaries, encrypted archives, minified bundles, oversized padded files, symlinks that point outside the skill.
Measured (Windows 11, Python 3.14):
| Target | Result | Time |
|---|---|---|
| Seeded attack skill (8 planted behaviours, inert) | DANGEROUS, all 8 reported | 0.04 s |
| anthropics/skills, 20 skills, 440 files, 11.7 MB | 14 SAFE, 6 REVIEW, 0 DANGEROUS | 5-16 s |
About half of those REVIEW lines were real behaviour, such as a local socket or a bundled archive. The rest were docs that quote risky phrases.
What it does not do: it is static analysis. A skill that downloads its real payload later shows up only as its network call. Binaries are flagged, not decompiled. It is a strong first filter, not a guarantee.
The audited skill cannot silence the scanner: there is no in-file ignore marker. Quoted excerpts are sanitized and labelled as untrusted data, so the report itself cannot inject your agent. Includes a GitHub Actions gate for shared skill repos.
Works in Claude Code, Cursor, Codex CLI, Copilot and any SKILL.md agent.
Privacy: runs fully locally, standard-library Python, no network, no telemetry.
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
4 uses
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 27 days ago
- Free with an Agensi account
Requires