The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality relies on ingesting large amounts of untrusted data from the internet via web_search and web_fetch. This data is then used to formulate research questions and drive the execution of sub-agents.
Ingestion points: External data enters the context through web_search (configured for a broad count of 20 results) and web_fetch (used for deep extraction from targeted pages).
Boundary markers: The instructions do not specify any delimiters or safety markers to isolate the fetched external content from the agent's internal instructions, creating a surface where malicious web content could influence the agent's reasoning or research direction.
Capability inventory: The agent has access to powerful tools including web_search, web_fetch, memory_search, and sessions_spawn, which can be leveraged to execute a multi-step research chain based on external input.
Sanitization: There is no evidence of content sanitization or filtering applied to the results of web operations before they are analyzed and used to generate further actions.
[DYNAMIC_EXECUTION]: The skill utilizes the sessions_spawn tool to delegate research tasks to sub-agents in Phase 3. The tasks for these sub-agents are generated dynamically based on the findings and "gaps" identified from untrusted web sources, which could lead to the propagation of malicious instructions into the sub-agents' execution context.

academic-deep-research