The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code files from user repositories, exposing it to potential indirect prompt injection attacks if analyzed code contains adversarial instructions. \n
Ingestion points: The skill reads local files using Read, Grep, and Glob tools when triggered by design pattern analysis keywords. \n
Boundary markers: Absent. The instructions do not define clear structural delimiters or prompt instructions to isolate user code content from the system guidelines. \n
Capability inventory: The skill's configuration includes the Bash tool, creating a potential exploitation vector if the model is successfully manipulated by malicious code comments. \n
Sanitization: Absent. No filtering, validation, or escaping mechanisms are applied to the code contents before processing.

codeprobe-patterns