Question 1

How does SkillzBench handle prompt injection risks in agent skills?

Accepted Answer

SkillzBench's Verify pillar runs security scans on instruction-only skills (SKILL.md plus text references) and pins every version to a commit SHA, so malicious instruction content is flagged before it reaches your agents. Script-bearing skills are listed link-out only and never distributed, which limits exposure to a harder class of threats while the platform is scoped to instruction packages.

Question 2

Can we enforce an allowlist of approved skills across our organization's AI agents?

Accepted Answer

Yes — the Teams tier is being designed for exactly this: organizational loadouts, allowlists and audit capabilities, on top of per-agent loadouts where each agent only sees the skills and tools explicitly toggled for it via the single MCP server.

Question 3

What does the Trust score actually measure, and can vendors pay to improve it?

Accepted Answer

The 0-100 Trust score is computed from security, quality, license, maintenance and community signals. Money never touches scores — paid tiers unlock more agents and unlimited enabled skills, but a creator cannot buy a better Trust score, and Lift is always shown separately from ratings.

Question 4

Is SkillzBench available now, and what does rollout look like for a security team?

Accepted Answer

SkillzBench is pre-launch and actively being built — the waitlist and newsletter are live today. Joining the waitlist puts your team in line for early access to the Discover, Verify, Benchmark, Equip and Stay-current pillars, so you can begin vetting skills as soon as the platform goes public.

Question 5

How is this different from just reviewing skills ourselves or relying on marketplace ratings?

Accepted Answer

SkillzBench is an independent testing lab, not a marketplace. Every skill is version-pinned to a commit SHA, validated, license-checked, security-scanned and quality-judged, then benchmarked in a sandbox with the same model with and without the skill — producing a measured Lift rather than a self-reported rating, with continuous re-scans when upstream changes land.

Agent Security Scanning for Enterprise Teams