The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies a surface area for indirect prompt injection as it is designed to ingest and process untrusted external data.
Ingestion points: External data and web content enter the agent's context through the BrowserOS tool mentioned in Section 4.
Boundary markers: The instructions explicitly state that external data must be "treated as hostile," which acts as a behavioral guideline for the agent.
Capability inventory: The skill maintains capabilities for terminal command execution via the ExecutionProxy Interface and uses the gemini-cli-security suite for offensive audits.
Sanitization: The protocol requires redacting secrets and PII (Personally Identifiable Information) before sharing context with other agents.
[COMMAND_EXECUTION]: The skill utilizes an ExecutionProxy Interface to perform terminal actions, such as running tests or audit commands (e.g., using the rtk prefix like rtk npm test). This allows the agent to interact directly with the underlying system to perform its security auditing tasks.

cybersecurity