reflect
Pass
Audited by Gen Agent Trust Hub on Aug 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill's primary workflows defined in
SKILL.mdrequire the execution of included Python scripts (scripts/calibration_scorer.pyandscripts/reflection_prompt_generator.py) to process user data. These scripts are deterministic, self-contained, and do not perform network or sensitive file operations.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from JSON files and uses it to generate analysis and reflection prompts. - Ingestion points: Data enters the system via JSON files loaded by
scripts/calibration_scorer.py(prediction logs) andscripts/reflection_prompt_generator.py(commitment logs). - Boundary markers: The skill lacks explicit boundary markers or instructions to the agent to ignore potentially malicious content within the processed logs.
- Capability inventory: The skill possesses the ability to execute shell commands via Python 3 to run its internal analysis scripts.
- Sanitization: While the scripts perform basic validation on numeric and status fields (e.g., confidence scores and commitment statuses), the text content of the predictions and commitments is processed without sanitization or escaping.
Audit Metadata