dreamina-edit-image
Fail
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- Indirect Prompt Injection (HIGH): The skill provides a mechanism for processing untrusted external content through the
promptparameter. - Ingestion Point: The
promptvariable in theedit_image_with_referencefunction. - Boundary Markers: None detected. The prompt is simply concatenated with
##inside the JSON payload. - Capability Inventory: Network write operations via
requests.postto external (jimeng.jianying.com) and internal (bytedance.net) endpoints. - Sanitization: No escaping or validation is performed on the user-supplied string before it is embedded in the
draft_contentJSON and transmitted. - Data Exposure (MEDIUM): The tool requires a
sessionid(cookie) for authentication. While not hardcoded, the logic handles sensitive credentials which could be exfiltrated if the agent's logic is subverted through the aforementioned prompt injection surface. - Command Execution (LOW): The skill utilizes Python's
requestslibrary to perform network operations anduuid/hashlibfor payload generation. While standard, these tools are used to interact with non-whitelisted external domains.
Recommendations
- AI detected serious security threats
Audit Metadata