natural-writing
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill mandates that the agent fetch and process content from two independent online sources for every writing task. This creates a vulnerability to indirect prompt injection, where an attacker-controlled website could contain hidden instructions that influence the agent's output or behavior when the agent reads the page for 'style research'.
- Ingestion points: The agent is instructed to 'open and read actual examples from at least two independent online sources' discovered via search results.
- Boundary markers: The instructions lack boundary markers or warnings to the agent to disregard any commands or instructions found within the fetched external content.
- Capability inventory: The skill assumes the agent has the capability to browse the web and fetch content from URLs to perform its research task.
- Sanitization: No sanitization or filtering of the fetched content is specified; the agent is explicitly told to let observations from these external sources 'inform the draft'.
Audit Metadata