The Agent Skills Directory

[DATA_EXFILTRATION]: Accesses local project configuration and source code for context grounding. The skill instructions in SKILL.md and references/search-prompts.md direct the agent to read package.json, local config files, and code to focus research on project-specific gaps. While useful for grounding, this involves exposure of local metadata and code structure to the agent context.- [EXTERNAL_DOWNLOADS]: Retrieves data from arbitrary external search engines and websites. The skill utilizes WebSearch and WebFetch tools to pull content from external URLs as its primary function. This behavior is expected given the skill's purpose.- [PROMPT_INJECTION]: Vulnerable to indirect prompt injection via ingested web content. * Ingestion points: Data is ingested from the open web using WebFetch and WebSearch tools within the search agent prompts (references/search-prompts.md). * Boundary markers: The instructions lack explicit delimiters or instructions to treat web-fetched content as untrusted data, which could lead the agent to follow malicious instructions embedded in a website. * Capability inventory: The skill has file-read access to local code and configurations, network-read access, and file-write access to the ~/.local/state/web-research/ directory. * Sanitization: No evidence of sanitization or content filtering for ingested web data is present before the synthesis phase.

web-research