video-frame-extractor
Fail
Audited by Gen Agent Trust Hub on Jul 27, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [CREDENTIALS_UNSAFE]: The skill contains a hardcoded API key for the Coze platform in both
SKILL.mdandscripts/coze_bot_client.py. The keycztei_qHZQ0A5OSJjsmfZWmVb8bqu2BTbtB240YGbDYLhZpsIr8jER4aL4Aevyii8rnKfNsis provided as a default value and in configuration examples, exposing it to unauthorized use. - [DATA_EXFILTRATION]: The scripts
scripts/coze_bot_client.pyandscripts/visual_analyzer.pyread local image files and transmit them to external endpoints, includinghttps://api.coze.cn/v3/chatand other user-configurable vision API providers like OpenAI. - [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection where malicious instructions embedded within video frames could influence the behavior of the vision model or downstream agents.
- Ingestion points: Visual content from video frames processed by AI models in
scripts/visual_analyzer.pyandscripts/coze_bot_client.py. - Boundary markers: Absent; the prompt templates provided in
references/analysis-guide.mddo not include instructions for the model to ignore text or instructions found within the images. - Capability inventory: The skill can write to the local file system (
scripts/video_frame_extractor.py) and perform network requests (scripts/visual_analyzer.py). - Sanitization: No sanitization or validation is performed on the data returned by the vision APIs before it is saved to structured JSON files.
Recommendations
- AI detected serious security threats
Audit Metadata