assemblyai-transcribe
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEMETADATA_POISONINGINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [METADATA_POISONING]: The skill's YAML frontmatter identifies the author as "OpenAI," which contradicts the provided author context of "tristanmanchester." This represents deceptive metadata that could lead users to misattribute the source of the skill.\n- [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted audio/video content, transcribes it, and passes the resulting text to an LLM via the
llmcommand. If an attacker embeds malicious instructions within the audio, the downstream LLM may execute them.\n - Ingestion points: Audio and video files processed in
scripts/assemblyai.mjsthrough thetranscribeandllmcommands.\n - Boundary markers: The logic for building LLM prompts in
buildLlmInputFromAgentJsonand thellmcommand uses basic labels like "Task:" and "Transcript:" but does not employ strong delimiters or explicit instructions to the LLM to ignore instructions embedded in the data.\n - Capability inventory: The skill possesses capabilities for file system writes (
--out,--bundle-dir) and network requests to external APIs.\n - Sanitization: There is no evidence of transcript content sanitization before interpolation into LLM prompts.\n- [EXTERNAL_DOWNLOADS]: The skill communicates with AssemblyAI's official endpoints (
api.assemblyai.com,llm-gateway.assemblyai.com) to perform transcription and analysis. These are recognized well-known services essential for the skill's operations.
Audit Metadata