voice-ingest

Pass

Audited by Gen Agent Trust Hub on Apr 29, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses Bash tools (file, test, mkdir) to interact with the filesystem. These tools are invoked with user-supplied arguments from the audio path input, which represents a surface for command injection if the execution environment does not perform strict sanitization.- [PROMPT_INJECTION]: The skill processes untrusted audio transcripts to classify content and suggest routing actions, creating a surface for indirect prompt injection.
  • Ingestion points: Untrusted transcript data is ingested via the claude-transcription companion skill in SKILL.md.
  • Boundary markers: The skill uses Markdown headers and frontmatter blocks to structure the data, though it lacks specific instructions to ignore embedded commands within the transcript.
  • Capability inventory: The skill possesses Write access to the local filesystem and the ability to execute restricted Bash commands.
  • Sanitization: The instructions explicitly forbid paraphrasing or cleaning the transcript (Guardrails section), meaning malicious instructions in the audio will be preserved verbatim in the stored inbox files.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 29, 2026, 05:16 AM
Security Audit — agent-trust-hub — voice-ingest