The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection because it processes untrusted data from execution transcripts and output files.\n
Ingestion points: Data is ingested from transcript_path and all files within outputs_dir using the Read and Glob tools.\n
Boundary markers: No delimiters or specific instructions are provided to distinguish between the agent's instructions and the untrusted content being evaluated.\n
Capability inventory: The agent uses Read, Grep, Glob, and Write. Malicious instructions inside the evaluated files could potentially influence the agent's file writing or grading decisions.\n
Sanitization: Content from external files is processed and searched without sanitization or validation.\n- [DATA_EXFILTRATION]: The skill presents risks related to sensitive file exposure and unsafe file operations.\n
The use of user-defined paths for transcript_path and outputs_dir allows the agent to read arbitrary files, which could lead to the exposure of sensitive configuration or credential files if those paths are provided as input.\n
Step 5 performing a write to outputs_dir + "/../grading.json" constitutes a directory traversal vulnerability, as an attacker could manipulate the directory path to write the output file to unintended locations on the filesystem.

skill-grader