The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch and synthesize data from live web sources, which inherently introduces the risk of the agent encountering and obeying malicious instructions embedded in external web pages.
Ingestion points: The fetch_webpage tool is used to retrieve content from external URLs as described in the Workflow section (SKILL.md).
Boundary markers: The instructions do not define specific delimiters or "ignore instructions" to separate fetched untrusted data from the agent's core instructions.
Capability inventory: The agent is equipped with fetch_webpage for reading external data and is encouraged to use figma-doc for generating documents based on that data.
Sanitization: The synthesis and quality checklist steps focus on accuracy and citations but lack requirements for sanitizing or filtering potential injection patterns or control characters from fetched content.
Mitigation: To reduce risk, the agent should wrap external content in clear delimiters and be instructed to ignore any instructions found within the fetched data.

online-research