seedance-2-5
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill reads local files specified by the user and transmits their content to the external API endpoint (ai-factory.frevana.com).\n
- Evidence: In scripts/generate_video.sh, the format_image_source function reads file data, base64-encodes it, and includes it in the JSON payload sent via curl to the Frevana video endpoint.\n- [COMMAND_EXECUTION]: The skill relies on shell script orchestration that executes embedded Python snippets to handle complex data processing tasks.\n
- Evidence: scripts/generate_video.sh contains multiple Python heredocs used for base64 encoding image files and constructing JSON payloads for API requests.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from the user which is then passed to a video generation model, creating a potential surface for indirect prompt injection.\n
- Ingestion points: Found in scripts/generate_video.sh via the --prompt, --prompt-file, --first-frame, and --last-frame parameters.\n
- Boundary markers: The skill does not implement delimiters or specific instructions to ignore embedded commands within the text prompts.\n
- Capability inventory: The skill has the ability to read local files and perform network operations via curl.\n
- Sanitization: While the skill uses json.dump to ensure the integrity of the JSON payload structure, it does not sanitize the prompt text or validate the content of the image files.
Audit Metadata