The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on an indirect prompt injection surface because it ingests untrusted data from external web pages and user-provided browser recordings, while maintaining powerful browser interaction capabilities.
Ingestion points: Processes untrusted external data from websites via snapshot -i and get text body commands, as well as third-party Chrome DevTools Recorder files (JSON/JS) in SKILL.md.
Boundary markers: Absent; there are no specific delimiters or security constraints instructing the model to ignore potential malicious prompts embedded within the webpage text or recordings.
Capability inventory: Possesses browser control capabilities including form filling, click execution, navigation, and spawning sequential subagents for multi-platform automation.
Sanitization: Absent; webpage text and recording data are parsed directly without validation or sanitization routines.
[COMMAND_EXECUTION]: Executes shell instructions via the agent-browser CLI to manage browser states, list tabs, and perform interactive elements automation. This capability is tied directly to the primary, expected function of the skill.
[DYNAMIC_EXECUTION]: Utilizes agent-browser --auto-connect eval to dynamically execute JavaScript snippets. This execution occurs within the browser webpage sandbox environment to traverse complex page elements or same-origin iframes.

chrome-automation