alterlab-cdm-documentary-research
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructions facilitate the ingestion of untrusted data which could contain adversarial instructions intended to influence the agent's behavior.
- Ingestion points: The agent is explicitly instructed in the 'Workflow' and 'Agentic Protocol' sections of
SKILL.mdto "Search the web for archival sources" and "Read existing project files for context — research notes, preliminary treatments, subject lists, or production logs." - Boundary markers: The skill does not define any boundary markers or instructions to the LLM to ignore potentially malicious instructions embedded in the retrieved web content or user files.
- Capability inventory: The skill possesses the capabilities to perform web searches and write files to the project directory (e.g.,
{project}-treatment.md,{project}-interview-questions.md). An attacker could potentially use indirect injection to cause the agent to overwrite files or perform specific search queries. - Sanitization: There are no sanitization or validation protocols described to filter content retrieved from external sources before the agent processes and acts upon it.
Audit Metadata