Breaking

Cisco Releases Open-Source Scanner to Detect Threats in AI Agent Skills

June 30, 2026 at 01:33 EDT

Security
AI Agents
Open Source

Cisco's AI Defense team has released skill-scanner, an open-source security scanner on GitHub that detects prompt injection, data exfiltration and malicious code hidden in AI agent "Skills." As the reusable instruction sets that agents draw on become a new attack surface, the tool—aimed at pre-install inspection—has drawn attention among developers.

Open Source · Cisco AI Defense

Cisco Open-Sources "skill-scanner" to Hunt Prompt Injection in AI Agent Skills

A static security scanner that analyzes SKILL.md instruction files for prompt injection and data exfiltration — as malicious third-party "skills" become a new supply-chain attack surface.

36%

of public skills found vulnerable in "ToxicSkills" research

1,467

vulnerable skills identified across public repositories

detection engines combined in one local scan

Engine depth: Cisco vs NVIDIA

Detection engines stacked per tool — taller column = more layers of analysis

Static

Behavioral

LLM

Cloud

What it scans for

Purpose vs behavior gap

declared intent ≠ actual actions

Obfuscation

Unicode smuggling, base64

Hidden directives

"ignore previous instructions"

Runs locally (pip install cisco-ai-skill-scanner, Python 3.10+) without sending source externally by default · outputs SARIF, HTML, JSON, Markdown, CSV · GitHub Actions, pre-commit & VS Code support.

Welcomed

"The kind of tool that should exist" — even overdue. As skills proliferate, "every framework is about to look very exposed."

The limits

Static scans catch obvious patterns, not runtime behavior or two clean skills that chain. Bypasses are already reported — "trust but verify."

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.