03-character-consistency-technique
Pass
Audited by Gen Agent Trust Hub on Aug 12, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read character identity records from 'characters/.md'. This constitutes an ingestion point for data that could potentially contain malicious instructions if the records are influenced by untrusted sources. * Ingestion points: 'characters/.md' identity records accessed in Section 2. * Boundary markers: Absent; there are no instructions for the agent to use delimiters or ignore embedded instructions within these records. * Capability inventory: The skill involves performing API calls to external video generation providers (HeyGen, Google Flow) as described in Sections 2 and 3. * Sanitization: No explicit sanitization, validation, or escaping of the identity record content is mentioned before it is processed.
- [DYNAMIC_EXECUTION]: The skill includes shell ('run-eval.sh') and Python ('contract_check.py') scripts in the 'eval/' directory. While these are for development validation, they represent executable code bundled with the skill. * Evidence: 'eval/run-eval.sh' executes 'python3' with a heredoc script and local JSON input.
Audit Metadata