screenkite-transcription-cut
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The
scripts/apply_cuts.pyscript executes the ScreenKite CLI usingsubprocess.run. The implementation uses a list of arguments, which prevents shell command injection, and all input parameters are properly serialized as JSON strings before execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes transcript data from an external transcription service, creating a surface for indirect prompt injection.\n - Ingestion points: The script
scripts/compute_silence_cuts.pyreads a word-level JSON transcript to identify silence intervals.\n - Boundary markers: The
SKILL.mdworkflow mandates a human-in-the-loop review (Phase 4) and explicit user confirmation before any destructive changes are made to the project timeline.\n - Capability inventory: The skill has the capability to modify project state and duration by executing the ScreenKite CLI binary via the
apply_cuts.pyscript.\n - Sanitization: The processing logic in
scripts/compute_silence_cuts.pyis restricted to extracting numeric 'start' and 'end' timestamps, effectively ignoring any textual content that might contain malicious instructions.
Audit Metadata