About this skill

Problem

Agent skills may bundle prompts, code snippets, declared permissions, and network dependencies. Installing them without review can let risks surface only at runtime: prompt injection, credential leakage, dangerous execution, sensitive file access, unexpected outbound requests, or memory pollution. SkillGuard performs a pre-install security check that can be reviewed before adoption.

How It Works

It maps eight detectors to OWASP ASI Top 10:
- Prompt injection: flags jailbreak patterns, manipulative phrasing, and Chinese bypass attempts.
- Secrets and execution: checks for API keys, tokens, passwords, and dangerous calls such as eval, exec, and subprocess.
- Dependencies and permissions: audits dependency packages and cross-checks declared permissions against actual behavior.
- Data and network: detects sensitive file access, data exfiltration paths, and URL-based network requests against a whitelist.
- Memory risk: identifies memory poisoning and cognitive attack indicators.

After the scan, it outputs a TRACE score across Trust, Reliability, Authenticity, Compliance, and Exposure, making it easier to decide whether to publish, review, or block a skill.

Boundaries

SkillGuard is best used as a static pre-install review and gating check. It should not replace sandboxed execution, human code review, or runtime network isolation. Scores and detector findings are risk signals; final decisions should consider business sensitivity and the Agent's permission boundary.

Screenshots

Use Cases

Run a static pre-install scan of a third-party Agent skill and record TRACE scores before adding it to a workspace.
Check candidate skills for prompt injection, jailbreak phrasing, and Chinese inducement patterns during intake review.
Inspect skill packages for exposed API keys, tokens, passwords, and dangerous calls such as eval, exec, and subprocess.
Cross-check declared permissions against actual behavior, sensitive file access, and URL whitelist network requests before publishing.

Best For

Engineers responsible for Agent workspace intake: need to block high-risk third-party skills before release.
Platform security reviewers: need reviewable scan results organized by OWASP ASI risk items.
Maintainers of private skill repositories: need to compare candidate skills using TRACE scores before publishing.
Team leads integrating external Agent capabilities: need prompt injection and credential leakage as launch gates.