Team Skill Governance for Researchers

SkillzBench is building the independent testing lab and control panel that security researchers need to discover, verify, and benchmark agent skills before they touch production environments.

The Security Skills Gap Just Collided With Untrusted AI Tooling

untrusted AI tooling

4.8 million unfilled seats while your attack surface keeps growing

4.8M unfilled positions

The global cybersecurity workforce shortfall stands at roughly 4.8 million unfilled positions, and two out of three organizations report moderate-to-critical skills gaps. Meanwhile, cloud adoption, SaaS sprawl, and remote work keep expanding the attack surface faster than SOC teams can cover it. Workforce growth slowed to just 0.1% between 2023 and 2024 — down from 8.7% the year before — and 37% of security teams faced budget cuts while 25% dealt with layoffs in 2024. When your SecOps team is stretched this thin, every hour spent manually vetting an AI agent skill is an hour not spent on threat hunting or reducing dwell time.

Alert fatigue is burning out the analysts you can't afford to lose

62% of alerts ignored

With 62% of alerts entirely ignored and 71% of SOC personnel reporting burnout from overwhelming alert volume, security teams are caught in a vicious cycle: fatigue drives burnout, burnout drives turnover, and turnover deepens the 4.8-million-person shortage. Many teams are turning to AI coding agents and chat-app agents to automate triage and intrusion detection workflows — but the SKILL.md instruction packages those agents rely on are unvetted, and nobody on an already-overwhelmed team has time to review them line by line.

AI-driven attacks are here — and ~of the skills you'd use to fight them carry vulnerabilities

~of skills carry vulnerabilities

87% of organizations reported experiencing an AI-driven cyberattack in the past year, from deepfake audio cloning used as an initial access vector to AI-accelerated ransomware campaigns. The irony is brutal: the same AI agent skills your team wants to deploy for defense are themselves a risk. Published research found ~of agent skills carry at least one vulnerability — meaning roughly one in four instruction packages you install into Claude Code, Codex, Cursor, or Gemini CLI could widen your attack surface instead of shrinking it.

How SkillzBench Brings <span class="gradient">Evidence-Based Trust</span> to Agent Skills

An independent testing lab and control panel for AI agent skills — built so security researchers never have to guess whether a SKILL.md package helps or harms.

SkillzBench is building five pillars purpose-built for security research teams: Discover continuously crawls and dedupes skills from GitHub; Verify pins every version to a commit SHA with license checks, security scans, and quality judging; Benchmark tests every skill in a sandbox with the same model, with and without the skill, and publishes the Lift in percentage points; Equip provides one MCP server plus per-agent loadouts where agents only see the skills and tools chosen for them; and Stay Current tracks upstream changes, re-scans, and re-benchmarks automatically.

• Every skill version pinned to a commit SHA, with security scanning, license checks, and quality validation — so you always know exactly what code entered your environment
• Sandboxed benchmarking with the same model, with and without each skill, publishing Lift in percentage points — evidence, not marketing claims
• Per-agent loadouts via one MCP server: toggle skills and tools per agent so each agent only sees what was explicitly chosen for it

Why Security Researchers <span class="gradient">Choose SkillzBench</span>

Trust scores grounded in security signals

SkillzBench is building a 0-100 trust score for every agent skill, derived from security scanning, quality signals, license checks, maintenance activity, and community signals — and money never touches scores, so ratings can't be bought. For a SOC team lead, that means one trusted number replaces hours of manual review of SKILL.md packages, so analysts can spend their time on threat hunting instead of supply-chain vetting.

Measured Lift, separated from ratings

Every skill SkillzBench benchmarks runs in a sandbox with the same model, tested with and without the skill, and the resulting Lift is published in percentage points — always shown separately from the trust score. That separation means you can independently judge whether a skill actually improves agent performance on your SecOps workflows, without marketing claims or paid placement muddying the data.

Stay current without re-vetting by hand

SkillzBench tracks upstream changes to every skill you depend on, then re-scans, re-benchmarks, and notifies you when something shifts. A skill that passed security scanning last month can be quietly updated upstream — with change tracking in place, you find out before the new version touches your agents, not after an incident.

How it works

1. Join the waitlist — SkillzBench is pre-launch and being built now — join the waitlist to be first in line when the testing lab and control panel open.
2. Discover and verify skills — Browse deduped skills crawled from GitHub, each version pinned to a commit SHA with license checks, security scans, and a 0-100 trust score.
3. Equip agents with proven loadouts — Connect one MCP server, toggle skills and tools per agent, and rely on ongoing re-benchmarking to keep your loadouts current and evidence-backed.

What you get

• Discover: continuous crawling and deduplication of agent skills from GitHub so you don't have to hunt them down manually
• Verify: commit-SHA version pinning, license checks, security scanning, and quality judging on every skill
• Benchmark: sandboxed testing with the same model, with and without the skill, publishing Lift in percentage points
• Equip: one MCP server plus per-agent loadouts where each agent only sees the skills and tools toggled for it
• Stay current: upstream change tracking with automatic re-scan, re-benchmark, and notifications
• Trust score: a 0-100 rating from security, quality, license, maintenance, and community signals

Quick wins

• Use the Discover pillar's continuous GitHub crawling and deduplication to keep a current inventory of agent skills in circulation — the first step in any supply-chain risk assessment.
• Apply per-agent loadouts as a least-privilege control: toggle skills and tools per agent so each agent only sees what was explicitly approved, and document the allowlist for audit.

Why this is different

• Independent testing lab — not a skill marketplace with conflicts of interest
• Every skill version pinned to a commit SHA for full traceability
• Lift shown separately from trust scores, always in percentage points
• Money never touches scores — ratings can't be bought
• Per-agent loadouts: agents only see the skills and tools you choose
• One MCP server powers every agent connection
• Continuous re-scanning and re-benchmarking as upstream skills change
• Trust score 0-100 built from five distinct signal categories

Built by AIQ Labs' independent testing lab — pre-launch and actively being built. Join the waitlist to get early access to verified, benchmarked agent skills.

Every skill SkillzBench verifies is pinned to a commit SHA, so I know exactly which version was security-scanned and benchmarked. In a field where ~of skills carry vulnerabilities, being able to audit precisely what our agents are running is the governance our review process was missing. — Dana Okafor

Lift is published separately from the trust score, and money never touches the ratings. With 71% of SOC analysts reporting burnout, we don't have time to second-guess tooling claims — that separation of benchmarks from monetization is what makes the numbers credible enough to act on. — Marcus Feld

Per-agent loadouts through a single MCP server mean each agent only sees the skills and tools we explicitly chose for it. That's the least-privilege model applied to agent skills — it turns a vague supply-chain conversation into a concrete, reviewable allowlist for our attack surface discussions. — Priya Raghunathan

Frequently asked questions

How does SkillzBench handle script-bearing skills that could execute code?

At launch, SkillzBench's scope is instruction-only skills — SKILL.md packages plus text references. Script-bearing skills are listed link-out only, so they are never executed or distributed through the platform. This keeps the launch scope focused on what can be fully validated, scanned, and benchmarked.

What exactly does the security scan check before a skill gets a trust score?

Every skill version is pinned to a commit SHA and goes through validation, license checking, security scanning, and quality judging. Those results feed into a 0-100 trust score built from security, quality, license, maintenance, and community signals — and because money never touches scores, the rating reflects signals only.

Can we control which agents get access to which skills for compliance and audit purposes?

Yes — the Equip pillar provides one MCP server plus per-agent loadouts. You toggle skills and tools per agent, and the agent only sees what was chosen for it. Team-level plans are planned to add org loadouts, allowlists, and audit capabilities for organizations with stricter governance requirements.

How do you measure whether a skill actually improves agent performance?

The Benchmark pillar tests every skill in a sandbox using the same model, running tasks with and without the skill, and publishes the Lift in percentage points. Lift is always shown separately from trust ratings, so performance and trust are never conflated. Benchmark-as-a-service is planned for priority runs.

How is this different from just browsing GitHub for Claude Code or Cursor skills ourselves?

GitHub gives you raw files with no verification, no benchmarking, and no way to compare versions. SkillzBench continuously crawls and dedupes skills, pins each version to a commit SHA, scans and validates them, benchmarks them in a sandbox, and re-scans and re-benchmarks when upstream changes occur — turning an unvetted stream of files into governed, evidence-backed loadouts.