SkillzBench launches free directory that security-scans and benchmarks AI agent skills before agents use them
Just another WordPress site
Press Release Disclaimer: This is a press release distributed through the XPR Media network. It has not been independently verified by our newsroom.
Halifax, Nova Scotia — October 23rd, 2026 — AIQ Labs today launched SkillzBench, a free, independent directory of AI agent skills that security-scans every skill, benchmarks it against the same AI model working without it, and publishes the result. A single connector then gives each agent only the skills its owner selects. The platform is live at SkillzBench.com.
Agent skills are packages of instructions, usually a SKILL.md file, that teach AI agents such as Claude Code, Codex, Cursor and Gemini CLI how to handle a specific kind of work. Thousands are now shared publicly on GitHub, with little vetting. A January 2026 study, “Agent Skills in the Wild” by Yi Liu and co-authors, analyzed 31,132 public skills and found that 26.1 percent contained at least one security vulnerability, and that about 5.2 percent showed patterns suggesting malicious intent.
“Anyone can publish a skill, and until now nobody could say whether a given one helps an agent, hurts it, or does something it shouldn’t,” said Pat Riley, founder of AIQ Labs. “SkillzBench runs the test and publishes the number, including when the number is negative. Then it lets people hand their agent exactly the skills that earned their place.”
How the platform works
SkillzBench continuously crawls GitHub for agent skills, groups them into categories and removes duplicate forks. Each skill version is pinned to a specific commit, so the version that was scanned is the version an agent receives. Every version is license-checked, security-scanned and reviewed by two independent AI judges before it is published. As of October 21, the platform had discovered 1,285 skills across 53 source repositories, run 1,376 security scans and published 799 skills. Skills that receive a do-not-install verdict are kept out of the directory. Each published skill carries a Trust Score from 0 to 100 built from security, quality, license, maintenance and community signals, with the weights published on the site.
How SkillzBench benchmarks skills
Each benchmarked skill is tested in an isolated sandbox with no internet access. The same AI model attempts the same tasks with the same tools, once without the skill and once with it. The difference in pass rate, called Lift, is published in percentage points with a 95 percent confidence interval. Results that show a skill making an agent worse are published alongside the gains. Some tasks are held out and never shown publicly, so a skill cannot be tuned to the visible tests, and deliverables are scored by a separate judge that never sees which version produced them. Promising skills are re-tested on additional models from Anthropic and OpenAI. To date the platform has run 14,663 benchmark trials, and 136 skills have a measured result. Full methodology is published at https://skillzbench.com/methodology.
Connecting an agent to its skills
Users connect an agent once through the SkillzBench MCP server, an open standard that AI apps use to reach outside tools. In the Claude desktop app and on claude.ai, the connection is a pasted link and a sign-in, with no key to copy. Developers can connect Claude Code, Codex, Cursor, Gemini CLI or any MCP-compatible client with a single command. Each connected agent has its own loadout, and users switch skills on or off per agent from a dashboard. An agent sees only the skills switched on for it and loads each one only when a task calls for it, which keeps its working memory free of instructions it does not need. One-click packs group proven skills for common work such as writing documents, building web pages, research, coding and security reviews. Setup instructions are at SkillzBench
Free, and supported by advertising
Every part of SkillzBench is free, including the directory, benchmark results, the connector, agent loadouts, packs and update alerts. The platform is funded by clearly labeled advertising in and around its articles and newsletter. Advertisers and sponsors cannot change a skill’s score, ranking, scan result or benchmark outcome.
Availability
SkillzBench is available now at SkillzBench.com. Skill authors can submit a GitHub repository for scanning and benchmarking at https://skillzbench.com/submit.
About SkillzBench
SkillzBench is an independent testing lab and control panel for AI agent skills. It discovers, verifies and benchmarks skills from GitHub and gives users one MCP connector to equip each of their agents with the skills they choose. SkillzBench is free and is not affiliated with Anthropic, OpenAI, Cursor or Google. Learn more at https://skillzbench.com.
About AIQ Labs
AIQ Labs is a Nova Scotia-based company that builds AI products and custom AI systems for businesses. More at https://aiqlabs.ai.
AIQ Labs launches SkillzBench, a free directory that security-scans AI agent skills, tests each against the same model without it and publishes the difference, plus one MCP connector that gives each agent only the skills its owner picks.
Media Contact
Company Name: SkillzBench
Contact Person: Pat Riley
Email: Send Email
Country: Canada
Website: https://skillzbench.com
Press Release Distributed by ABNewswire.com
To view the original version on ABNewswire visit: SkillzBench launches free directory that security-scans and benchmarks AI agent skills before agents use them
Media gallery
