SkillSpector: check an AI agent skill before you install it
NVIDIA's free, open-source scanner looks inside an AI agent skill for hidden instructions, data theft and permission tricks. What it does, its limits, and what we could not confirm.
Type
Security scanner for AI agent skills (command line)
License and cost
Apache-2.0 licence, free
Reviewed
29 Oct 2026
Version
Not stated in the README (no published releases)
← All tool profilesNo. 1 / 11
In one sentence
SkillSpector checks a skill, which is a folder of instructions and scripts you install in an AI agent such as Claude Code or Codex, before you install it. It warns about common traps: stealing data, skipping permissions or slipping in hidden orders (prompt injectionPrompt injection Hiding instructions in text, a web page or a file so that an AI assistant reads them and obeys them as if they came from you. Tap again to open the glossary).
Who it is for
Anyone who installs skills or MCP servers made by other people into an AI agent: founders, indie makers and small teams.
What the numbers say
The README cites NVIDIA's own research on agent skills. Out of 31,132 skills analysed, 26.1% had at least one vulnerabilityVulnerability A weakness in software, a setting or a process that an attacker could use to get in or do harm. Tap again to open the glossary and 5.2% showed likely malicious intent. We have not read the paper behind these figures.
The good
- Free and open source, from NVIDIA, as part of its verified-skills work.
- It accepts a folder, a single SKILL.md file, a Git link or a zip, and prints to the terminal or as JSON, Markdown or SARIF.
- Fast static analysis, plus an optional review by an AI model. With the
--no-llmoption the skill's contents stay on your machine. - According to the README it gives a risk score from 0 to 100: 0 to 20 low, 21 to 50 caution, above 50 "do not install".
The limits it admits
- It only reads code: it does not run the skill, so it cannot see what it really does.
- It can miss text in other languages, text inside images, and compiled or encrypted code.
- NVIDIA calls it "defense in depth, not a sandboxSandbox A fenced-off area where a program runs with limited access, so if it misbehaves it cannot touch the rest of the system. Tap again to open the glossary": it does not protect you if you install the skill anyway.
- NVIDIA's guide does not discuss false alarms or how complete the detection is, so a clean report does not prove a skill is safe.
When not to use it
- As the only barrier before giving an agent broad access.
- On skills you cannot read, or code that is not plain text.
- If you do not want the skill's contents sent to a model provider: use
--no-llm. Even then it checks dependency names and versions against OSV.dev.
What it costs you
It needs Python 3.12 or newer (Docker is optional). The AI review needs a key from a model provider or the login of a command-line AI tool you already use; the static mode needs neither. You update it as the repository changes.
Typical setup mistakes
- Sending a skill to an AI provider without checking where its contents go.
- Trusting a low score and giving the skill more access than it needs.
- Installing from a Git link without pinning a version. Pin a commit or tag if you want repeatable results (our advice, not NVIDIA's).
How to try it
We have not tested it, so we do not give our own step-by-step. NVIDIA's guide shows the command skillspector scan ./my-skill/ --no-llm and explains the options. According to the documentation, the command ends with exit code 0 if the score is 50 or lower, 1 if it is higher, and 2 if there was an error. Read the report and do not rely on the score alone.
What we could not confirm
We did not install or run the tool, and we did not read the research paper the README cites. The README gives no version or release date. The README says 71 detection patterns and a numeric risk score, while NVIDIA's guide says 68 patterns and does not describe the numeric score, so we do not rely on either number.
Related
Glossary: prompt injection, supply chainSupply chain attack An attack that reaches you through something you already trust and install, such as a code package, a tool or a provider. Tap again to open the glossary. Our own free checks live in Tools.
More free tools
Other profiles in this section.
- CrowdSec: a free, open source engine that reads your logs, spots attackers and blocks them, with a shared community blocklist: Intrusion detection and blocking engine (reads logs, blocks abusive IPs)
- Lynis: a free audit that checks a Linux server's security settings and tells you what to fix: Security audit and hardening tool (runs on the host)
- Semgrep: scan your own code for bugs and security mistakes before you ship, free and open source: Static code analysis scanner (command line)
- Bitwarden: an open-source password manager your team can self-host, free and paid options: Password manager (hosted service and self-hostable server)
- OpenVAS: a free scanner that checks your servers and apps for known weaknesses: Vulnerability scanner (part of the Greenbone Community Edition)
- OSV-Scanner: check the libraries your project uses for known vulnerabilities, free and open source: Dependency vulnerability scanner (command line)
- Trivy: scan your containers, code folders and dependencies for known vulnerabilities, free and open source: Vulnerability and misconfiguration scanner (command line)
- Wazuh: an open-source platform that watches your servers for suspicious activity: Security monitoring platform (agents, server, indexer and dashboard)
- Gitleaks: find keys and passwords committed to your code by mistake, free and open source: Secret scanner (command line)
- Restic: encrypted backups you can restore, free and open source: Backup program (command line)
Sources
Follow us
New profiles and security news: YouTube (@safet2ship) and Instagram (@safet2ship).