SuperSkill gives your company one private, vetted catalog of AI skills — your internal tools and the public ones you trust — served to each team with role-based access. Every skill is checked before it's published, so nothing from an unknown source reaches your agents.
Currently in beta.
Reviewed and published to the catalog — ready for every team.
A trusted public skill, vetted and added to your catalog.
Hidden instructions told the agent to ignore its rules and leak the conversation.
Read local API keys and .env files, then called out to an unknown host.
Shipped usage and file contents to a third-party endpoint no one approved.
Pulls in a package already flagged as malware.
Self-hosted, built on the open MCP skills standard, and ready for every agent you run — the catalog is yours.
Add your internal tools — your ERP assistant, your deploy scripts — alongside the public skills you trust. One source of truth for the whole company.
Before a skill is published, SuperSkill scans it and runs it in a sealed sandbox. A skill that misbehaves never makes it in.
Request a skill, and once it's cleared it's in the catalog — the agent pulls it straight from there, with no GitHub and no unvetted code.
Five independent checks run on every skill before it reaches the catalog.
Reads the code for risky behavior and the capabilities it asks for.
Scans every package it pulls in against known-bad advisories.
Matches live indicators of attacks seen in the wild.
Detonates it in an isolated microVM and watches what it does.
Optional model pass for intent and hidden instructions.
Every finding is mapped to CWE, MITRE ATLAS and OWASP — so you get one clear verdict, with the full report a click away.
A skill is instructions and code your agent will trust — which is exactly what makes it the new attack surface. What SuperSkill catches:
security checks covered in a 104-mechanism comparison of 11 skill scanners. The next best covers 46.
Full coverage counts 1, partial counts ½. The SuperSkill count is our own, taken from our code and tests; the other scanners' counts are the comparison author's. Others still lead us on CI and reporting integrations, scanner self-testing, and offline dependency data.
In the public MalSkillBench benchmark (4,000 safe and 3,945 malicious skills), SuperSkill with its AI review catches 84% of the malicious skills while wrongly flagging 11% of the safe ones. That puts it in the Top-2 of the scanners measured, ahead of every other tool and behind only one LLM-based scanner that wrongly flags twice as many safe skills. With the AI turned off, the static checks alone still catch 65% at a 6% false-alarm rate, the best result among scanners that need no language model: no AI calls, no per-scan cost, the same answer every time.
Score = share of malicious skills caught minus share of safe skills wrongly flagged (higher is better; the benchmark authors' own scripts). “LLM” marks scanners that use a language model. The AI figure is the SuperSkill engine with a main model plus a second model that re-checks every flag; its safe-skill rate is estimated from a 1,000-skill sample, about one point of error. Other scanners' results are the benchmark's own. Measured 17 October 2026.
Built to run for an organization. Admins get a single console to curate the catalog, review and approve skills, sign and audit, and control who sees what.
Publish your internal skills and the public ones you trust. You decide exactly what's in the catalog your company serves.
People ask for a skill; you review and approve in a dashboard. In enforce mode, only approved skills are ever served.
OAuth sign-in with roles and skill groups decides which teams see — and who can publish — which skills.
Every served skill is signed, with provenance and a bill of materials, so you can prove exactly what shipped and from where.
A hash-chained log with an outside witness records every approval, override and revocation — it can't be quietly rewritten.
A published skill turns out bad? Open an incident and pull it from every team at once.
SuperSkill is in beta and sign-in is limited to approved accounts. Tell us who you are, what you want to use it for, and which skills you would like to test. We review every request and get back to you at the address you give.