The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides commands to access and retrieve sensitive data from the browser environment, including cookies (agent-browser cookies), localStorage, and sessionStorage (agent-browser storage). Furthermore, it includes a file upload capability (agent-browser upload) that allows the agent to read files from the host's local filesystem and transmit them to external websites, which could be leveraged for unauthorized data exfiltration.
[COMMAND_EXECUTION]: The skill supports the execution of arbitrary JavaScript within the browser's context via the eval command. Additionally, it allows the agent to specify a custom path for the browser binary using the --executable-path flag, providing a mechanism to execute arbitrary local binaries.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted data from the web without sufficient isolation or sanitization.
Ingestion points: agent-browser snapshot, agent-browser get text, agent-browser get html, and agent-browser get value across SKILL.md and references/advanced.md.
Boundary markers: Absent. There are no instructions for the agent to use delimiters or to disregard instructions contained within the browser's output.
Capability inventory: The skill possesses network access (open), file system read access (upload), file system write access (screenshot, pdf, state save), and dynamic code execution (eval).
Sanitization: Absent. The agent is expected to interpret and act directly upon the information retrieved from web pages.
[EXTERNAL_DOWNLOADS]: The skill's setup instructions involve downloading and installing the agent-browser tool from the public npm registry.

browser