deepgram-migration-deep-dive
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEDYNAMIC_CONTEXT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_CONTEXT_INJECTION]: The skill uses dynamic context markers in
SKILL.mdto execute shell commands at load time, specifically checking for the installation status of@deepgram/sdk,@aws-sdk/client-transcribe, andgoogle-cloud-speech. These are benign status checks. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process transcription data from external providers, which introduces a surface for indirect prompt injection if the transcribed audio contains malicious instructions.
- Ingestion points: The
transcribeUrlandtranscribeFilemethods in the adapter implementations process content from remote URLs and local files. - Boundary markers: No boundary markers or 'ignore' instructions are used when processing the transcription output.
- Capability inventory: The skill is granted
Bashtool access and file system access (Read,Write,Edit). - Sanitization: There is no evidence of sanitization or filtering of the transcript text before it is returned to the agent context.
Audit Metadata