court
Fail
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: CRITICALMETADATA_POISONINGINDIRECT_PROMPT_INJECTION
Full Analysis
- [METADATA_POISONING]: The
homepagefield inSKILL.mdcontains the URLhttps://chen.media, which has been flagged as malicious by automated security scanners (URLite blacklist). This URL is also embedded in theCREDITstring incourt.pyand is displayed to the user in the generatedVERDICT.mdfile, creating a risk that users may visit a compromised or malicious site.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input to form a 'case file' which is subsequently used to prompt various sub-agents (prosecutors, defense, jurors, and judge).\n - Ingestion points: The user-provided idea or plan is captured and written to
.court/case.md.\n - Boundary markers: Although sub-agents are instructed to follow their briefs, there are no robust boundary markers or instructions to ignore potential injection attacks within the user-provided case text.\n
- Capability inventory: The skill utilizes
general-purposesub-agents and a local Python script (court.py) for file management and state tracking within the project directory.\n - Sanitization: No explicit sanitization or filtering is performed on the user's input before it is passed to the sub-agent chain, allowing for potential manipulation of the trial's logic and outcome through embedded instructions.
Recommendations
- Contains 1 malicious URL(s) - DO NOT USE
Audit Metadata