The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill supports an eval command that allows the execution of arbitrary JavaScript within the browser context. It also allows for dynamic JS conditions in the wait --fn command. These features provide a direct path for executing arbitrary code inside the browser session.
[COMMAND_EXECUTION]: The skill instructions include global options for the agent-browser CLI, such as --executable-path and --extension. These flags permit the selection of arbitrary binaries to run as the browser or the loading of potentially malicious extensions from local paths.
[DATA_EXFILTRATION]: The skill includes functionality to extract and save sensitive browser data, including full session cookies and local storage contents (agent-browser cookies, agent-browser storage). While documented for session persistence, these capabilities can be used to harvest authentication tokens.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill reads untrusted data from external websites via commands like snapshot, get text, and get html (detected in SKILL.md).
Boundary markers: None identified in the instructions for separating website content from agent instructions.
Capability inventory: The skill has significant capabilities including network navigation, cookie retrieval, JavaScript execution, and local file uploads (detected in SKILL.md).
Sanitization: There are no explicit sanitization or filtering steps mentioned for the content retrieved from the web.

agent-browser