The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill ingests untrusted user data in the form of debate topics and questions, which are then used to generate multi-persona simulations. This creates an attack surface where malicious input could attempt to manipulate the 'committee' results.
Ingestion points: The CONVENE and PRESENT methods in CARD.yml accept user-defined topic and question strings.
Boundary markers: Absent. The skill does not explicitly define delimiters to separate user input from its internal persona instructions.
Capability inventory: The skill possesses the write_file capability (defined in CARD.yml), which it uses to record transcripts and decisions to the local filesystem.
Sanitization: No evidence of input validation or sanitization was found in the provided protocol files.

adversarial-committee