teach
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The SKILL.md instructions include a directive to 'open the lesson file for the user by running a CLI command.' This encourages the agent to use shell capabilities to interact with the host system, which presents a risk if the agent constructs commands using unvalidated file paths or metadata.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). The instructions mandate that the agent 'Never trust your parametric knowledge' and instead rely on external resources listed in RESOURCES.md. Content ingested from these external sources could contain adversarial instructions that influence the agent's lesson generation or workspace management.
- [PROMPT_INJECTION]: The framework uses MISSION.md and learning-records as primary context for steering the agent's behavior. These files ingest user-supplied data which then guides the agent's logic for selecting topics and generating files.
- [DATA_EXPOSURE]: The skill creates a NOTES.md scratchpad for recording 'user preferences, or working notes.' This file acts as a persistent repository for potentially sensitive contextual information about the user or their workspace, which the agent is instructed to refer back to frequently.
Audit Metadata