The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accepts a file path as input and directs the agent to read its contents without enforcing workspace restrictions or validating the file extension. This behavior could be exploited to access and expose sensitive information from system files or configuration files (e.g., SSH keys or environment variables) if an absolute or malicious path is provided.
[PROMPT_INJECTION]: The skill implements an indirect prompt injection surface by reading and analyzing untrusted data from external files (Ingestion point: ${input:filePath}). It lacks boundary markers to isolate the file content from the agent's primary instructions. Combined with the agent's capability to use tools like context7 and generate structured reports, malicious instructions embedded in the analyzed files could manipulate the recommendation logic or trigger unintended behavior. No sanitization or validation steps are defined for the ingested file body.

model-recommendation