The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing the agent-browser CLI globally via npm and provides steps to build the tool from source using a public repository from Vercel Labs.
[REMOTE_CODE_EXECUTION]: The agent-browser eval command allows for the dynamic execution of arbitrary JavaScript within the browser's context. This is a core feature for automation but poses a risk if the agent is manipulated into executing code provided by a malicious website.
[DATA_EXFILTRATION]: The skill includes commands to access and save browser session data, such as agent-browser cookies and agent-browser state save auth.json. These features handle sensitive session tokens and authentication data which could be targeted for exfiltration.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it allows an agent to ingest and act upon data from arbitrary external websites.
Ingestion points: External content is retrieved using agent-browser snapshot, agent-browser get text, and agent-browser get html across various commands in SKILL.md.
Boundary markers: Absent; the skill does not instruct the agent to use delimiters or specific safety instructions when processing retrieved web content.
Capability inventory: High; the agent can interact with pages (click, fill), execute script (eval), and manage authentication state (cookies, state).
Sanitization: Absent; no content filtering or validation mechanisms are described for data retrieved from web pages.
[CREDENTIALS_UNSAFE]: The command agent-browser set credentials is used to provide HTTP basic authentication details. While a standard feature, it involves the handling of plain-text credentials within the command execution flow.

Agent Browser