speech-aware-cutting
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external transcript data which could potentially contain malicious instructions intended for the agent during the review phase. \n
- Ingestion points: The scripts/review_cuts.py script reads JSON transcript files (transcript-*.json) provided by the user or external tools. \n
- Boundary markers: Transcript content is processed and presented to the agent as a 'heard script' for quality review. The instructions lack explicit delimiters or warnings to ignore commands that might be embedded within the spoken text. \n
- Capability inventory: The skill uses a standalone Python script for analysis and standard reasoning for review; it does not provide the agent with tools to execute transcript content directly. \n
- Sanitization: Transcript text is extracted and concatenated without filtering for prompt injection patterns. \n- [SAFE]: The provided Python script scripts/review_cuts.py is restricted to the Python standard library and performs only local file reading and data processing. \n- [SAFE]: The SKILL.md file contains explicit budget-management instructions (SPEND protocol) that require the agent to seek user confirmation or check stated limits before performing credit-consuming tasks, serving as a significant financial safety guardrail.
Audit Metadata