The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides a mechanism for the agent to ingest untrusted data from web pages, which serves as a surface for indirect prompt injection.\n
Ingestion points: Web content is retrieved through accessibility tree snapshots, HTML element extraction, and network resource logs in scripts/cdp.mjs.\n
Boundary markers: The skill does not implement delimiters or safety instructions to distinguish external web content from the agent's internal logic.\n
Capability inventory: The skill possesses high-impact capabilities including arbitrary JavaScript execution (Runtime.evaluate), automated typing, and navigation.\n
Sanitization: No data filtering or sanitization is performed on information fetched from the browser.\n- [COMMAND_EXECUTION]: The skill executes local system commands to manage background daemon processes.\n
Evidence: In scripts/cdp.mjs, the script uses child_process.spawn to run itself in a detached mode to maintain persistent WebSocket connections to the browser.\n- [REMOTE_CODE_EXECUTION]: The skill enables the execution of arbitrary code within the browser context.\n
Evidence: The eval and evalraw commands provide a direct interface for the agent to execute JavaScript expressions and raw DevTools Protocol methods in active tabs.

chrome-cdp