patent-exam-policy
Warn
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill implements a 'Skill Evolution' feature that allows it to modify instructions in other directories, specifically targeting 'skills/patent-disclosure/prompts/**'. While this process requires explicit user confirmation and is restricted to a whitelist defined in 'prompts/apply_after_confirm.md', it allows for the dynamic modification of agent logic based on external inputs.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external websites to influence agent behavior through prompt updates. 1. Ingestion points: External patent policy websites are accessed via 'research.md' using search and fetch tools. 2. Boundary markers: The 'guardrails.md' file establishes boundaries, prohibiting the use of instructions that attempt to deceive or bypass safety filters. 3. Capability inventory: The skill has the ability to write to output directories and perform write operations on specific prompt files within the workspace. 4. Sanitization: The skill mandates the use of verifiable URLs and structured formats to minimize the impact of malicious content in the source data.
- [COMMAND_EXECUTION]: The skill utilizes network tools like 'curl' and 'WebFetch' to retrieve policy documents from established government domains (e.g., 'cnipa.gov.cn'). These operations are aligned with the skill's stated function of legal policy research.
Audit Metadata