The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted external data, creating an entry point for indirect prompt injection.
Ingestion points: Source code files from the user project environment enter the agent context during security scans as indicated in SKILL.md.
Boundary markers: Absent. There are no specifications for structural boundaries or explicit rules instructing the agent to ignore hidden directives within processed code.
Capability inventory: The frontmatter of SKILL.md exposes powerful capabilities including Bash, Read, Grep, and Glob tools.
Sanitization: Absent. No sanitization or safety filters are applied to the analyzed file contents prior to token processing.

codeprobe-security