recoup-content-write-caption

Pass

Audited by Gen Agent Trust Hub on Jul 23, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted data from an artist's past social media posts to diarize a voice and generate new captions, creating an attack surface for indirect prompt injection. It also includes specific instructions to ensure transparency with the user.
  • Ingestion points: Historical social media post data retrieved from the vendor's API and local artist context files located in the workspace.
  • Boundary markers: The skill does not utilize explicit delimiters to isolate untrusted historical content, but relies on a diarization process to filter intent.
  • Capability inventory: The agent can execute shell commands, perform network requests via curl, and read/write to the local filesystem.
  • Sanitization: The skill incorporates an "anti-slop fingerprint" and verification checks to ensure output fidelity. Additionally, the "analyze gate" instructions prevent the agent from concealing failures or low-quality results from the user.
  • [COMMAND_EXECUTION]: The skill provides functional shell command templates for data processing and API interaction.
  • Reference files include bash scripts using curl, jq, sed, tr, and ls for workspace management and communication with the platform API.
  • Sourcing instructions suggest the use of yt-dlp for downloading audio assets from external platforms when local files are missing.
  • [DATA_EXFILTRATION]: The skill accesses sensitive authentication credentials and artist metadata stored in the environment and local workspace.
  • It utilizes the RECOUP_API_KEY and RECOUP_ACCESS_TOKEN environment variables for authentication.
  • It reads identity metadata from workspace files like RECOUP.md to resolve artist identifiers.
  • Network activity involving these credentials is directed exclusively to the vendor's official domain (api.recoupable.dev).
  • [EXTERNAL_DOWNLOADS]: The skill is configured to fetch necessary content and metadata from the vendor's API and other external sources.
  • Downloads artist profiles, engagement metrics, and track analysis from the platform's research and content endpoints.
  • Supports downloading audio files from sandbox repositories or via external utilities like yt-dlp as a fallback mechanism.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 23, 2026, 08:26 AM
Security Audit — agent-trust-hub — recoup-content-write-caption