The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is built entirely around executing shell commands via the agent-browser CLI. This includes extensive use of the Bash tool to perform browser actions such as clicking, filling forms, and navigating URLs.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of an external CLI tool (agent-browser) via npm and the download of Chromium browser binaries (agent-browser install). These are necessary for the skill's functionality but involve downloading and running external code.
[OBFUSCATION]: The agent-browser eval command supports a -b or --base64 flag designed to execute Base64-encoded JavaScript strings. While documented as a utility to avoid shell quoting issues, this mechanism allows for the execution of hidden code within the browser context.
Evidence in SKILL.md and references/commands.md shows examples of piping Base64-encoded strings to the evaluator.
[DATA_EXFILTRATION]: The skill provides commands to save and load browser state files (agent-browser state save auth.json), which contain sensitive information like session cookies and local storage data. While documentation advises against committing these files, their creation and potential exposure represent a data security risk.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection.
Ingestion points: Reads arbitrary content from the web via snapshot and get text commands.
Boundary markers: None provided in the instructions to distinguish between agent instructions and page content.
Capability inventory: The agent can perform file system writes (screenshots, PDFs, state files), execute JavaScript in the browser (eval), and perform network operations.
Sanitization: No evidence of sanitization or filtering of the content retrieved from web pages before it is processed by the agent.

agent-browser