The Agent Skills Directory

[COMMAND_EXECUTION]: The instructions direct the agent to use the Bash tool to analyze the codebase (e.g., using grep and find), which involves executing shell commands in the local environment.
[EXTERNAL_DOWNLOADS]: The skill utilizes WebSearch and WebFetch to download and ingest content from external web sources, reaching out to non-whitelisted domains during its research phase.
[PROMPT_INJECTION]: The skill possesses an inherent attack surface for indirect prompt injection as it processes and synthesizes content from external websites without boundary markers.
Ingestion points: Untrusted data enters the agent context via WebSearch and WebFetch (SKILL.md).
Boundary markers: The skill lacks instructions for using delimiters or explicit "ignore instructions" directives for content retrieved from external sources.
Capability inventory: The skill enables the use of Bash, Write, and WebFetch, which provide a significant capability set that could be targeted by instructions hidden in external data (SKILL.md).
Sanitization: No validation or sanitization steps are provided for external content before it is used in synthesis or stored in memory.

deep-research