viral-video-teardown-remake
Fail
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFEDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill instructs the agent to download a remote ZIP archive from cdn.beatra.ai and execute bundled Python scripts (authorize.py, mcp_client.py) which creates a runtime dependency on externally hosted code.
- [DYNAMIC_EXECUTION]: The package includes a self-update mechanism that replaces its own files with new versions from a remote server without user review; this feature is documented as being enabled by default in the provided version.
- [CREDENTIALS_UNSAFE]: The skill performs an OAuth flow and stores a long-lived bearer token in plain text at a fixed local path (~/.beatra/credentials.json).
- [DATA_EXFILTRATION]: The skill transmits telemetry including the host system name, agent platform environment variables (e.g., claude-code), and package usage metadata to vendor-controlled servers (api.beatra.ai, mcp.beatra.ai).
- [COMMAND_EXECUTION]: The instructions require the agent to run shell commands including curl, unzip, and python3 to download, extract, and configure the external package.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied videos and text prompts which creates a surface for indirect injection without specified sanitization or boundary markers.
- Ingestion points: User video files and text prompts (referenced in Security & Safety Notes).
- Boundary markers: None identified in the instruction markdown.
- Capability inventory: Subprocess execution (python3) and network operations (curl).
- Sanitization: No content filtering or sanitization steps are described for external media inputs.
Recommendations
- AI detected serious security threats
Audit Metadata