SkillSec - Security testing for AI agent skills
Full adoption. Zero vetting.
A company-wide rollout of Claude and Cursor turned into a security blind spot when employees started importing agent skills from unverified sources. We built SkillSec to close the gap, and then kept building until it became a standalone product.
From bold bet to blind spot
A mid-size company went all-in on Claude and Cursor: enterprise licences, full team rollout, immediate productivity gains. Within weeks, employees were building and collecting agent skills from GitHub, Reddit, and AI forums to push output even further. Nobody was vetting them. That's when the security team noticed unusual patterns and realized AI tool usage needed closer attention. They reached out to us, and we ran a rapid audit. As a first step, we identified the need to start monitoring the skills users were uploading, which had begun flooding in at scale.
Skills from everywhere. Trust from nowhere.
Agent skills from unknown sources were running inside the company's Claude and Cursor environments with access to internal context and APIs. With dozens of skills already in circulation, manual review was impossible. They needed an automated gate that any team member could use, without requiring a security background to operate it.
Scan before you run.
We built SkillSec as a targeted security checkpoint → paste a GitHub URL, select the skills to test, and a multi-agent system probes each one for prompt injection, data exfiltration, and SSRF vulnerabilities. Results come back as a severity-graded report in minutes. What started as a client solution became a standalone product.
Multi-agent testing
Specialised agents probe each skill independently, simulating adversarial inputs across multiple attack vectors simultaneously.
GitHub-native workflow
Submit a repo URL, select skill folders, get results. No installation, no codebase changes, private repos supported.
Severity-graded reports
CLEAN to CRITICAL classification with specific evidence and remediation steps - actionable without security expertise.
Commit-based deduplication
Results cached per commit SHA. Unchanged skills are never retested, only what changed gets scanned.
Adoption kept moving. The risk didn't.
The company went from zero visibility into what skills were running in their environment to a lightweight approval gate that any team member can use. Unknown third-party skills get scanned before adoption. Internal skills get tested before they're shared. The productivity push continues on verified ground.
- ✓ MVP delivered
- ✓ Full product built
- Live — standalone SaaS
See solution in action
Open solution →Want to See It in Action? Request Your Demo Access
Fill out the form and we'll send you a demo access token, valid for 48 hours, so you can explore the solution yourself.
Request Demo Access
We build custom AI systems for small and mid-size businesses - from working prototype to production, with a clear process and defined outcomes at every step. No generic tools, no long commitments before you've seen results.
Copyright PYXERO 2026 | All rights reserved.