respond-to-eval
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
Bashtool to executepdftotextandpandocfor document conversion. It constructs shell commands by interpolating user-provided file paths ($0), which can be a vector for command injection if input paths are not properly sanitized by the underlying platform. - [INDIRECT_PROMPT_INJECTION]: The skill processes student evaluation comments, which are untrusted external data. These comments are clustered and used to generate teaching improvement plans and edit targets, creating a risk that malicious instructions hidden within student feedback could influence the agent's behavior.
- Ingestion points: Evaluation files provided via the
$0argument (CSV, TSV, TXT, MD, PDF, DOCX). - Boundary markers: None identified in the prompt instructions for delimiting the student text.
- Capability inventory: The skill has access to
Bash,Write,Agent, andTasktools, allowing it to execute commands, write files, and delegate sub-tasks. - Sanitization: While the skill instructs the agent to anonymize quotes by stripping identifying details, it does not include instructions for sanitizing or escaping potential prompt injection attempts within the feedback.
Audit Metadata