skillify

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.95). In ADAPT mode, the runtime fetches outsider-authored public GitHub/agentskills/skills.sh SKILL.md content (e.g., via git clone or gh api ... contents/SKILL.md) and then reads it to rewrite/classify, so that free text is ingested into the agent’s LLM context.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The ADAPT mode explicitly fetches external skill sources at runtime (e.g., via git clone or the GitHub raw API such as https://github.com// and agentskills.io links like https://agentskills.io) and then ingests that fetched SKILL.md content to drive the agent's analysis and rewrite, which is a runtime external dependency that directly controls prompts/instructions.