teach
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from external sources via
RESOURCES.mdand user inputs to generate educational content, creating a vulnerability surface. - Ingestion points: External URLs and descriptions in
RESOURCES.md, as well as user queries provided during the teaching sessions in theSKILL.mdworkspace. - Boundary markers: The skill uses specific Markdown templates (e.g.,
MISSION-FORMAT.md,LEARNING-RECORD-FORMAT.md), but lacks explicit delimiters or warnings to the model to ignore embedded instructions within resource descriptions. - Capability inventory: The agent is authorized to write files to the local directory (
./lessons/,./learning-records/, etc.) and is instructed to execute CLI commands to open these files. - Sanitization: Filenames are partially sanitized using a "dash-case" requirement, but there is no instruction to sanitize or escape the content of the generated HTML lessons, which could contain malicious scripts.
- [COMMAND_EXECUTION]: The skill instructs the agent to perform actions that involve the local operating system's command line.
- Evidence:
SKILL.mdcontains the instruction: "If possible, open the lesson file for the user by running a CLI command." This involves executing a tool (likeopen,xdg-open, or a browser) against a file whose content is generated from untrusted external resource data, creating a potential boundary crossing if the content is not properly handled.
Audit Metadata