The Agent Skills Directory

[SAFE]: The skill provides instructions for using a browser automation tool from a trusted organization and includes comprehensive security guidance.\n- [COMMAND_EXECUTION]: The skill instructs the agent to use the agent-browser CLI. Potential security risks associated with CLI commands are mitigated by a risk-categorized command reference in references/safety.md, which recommends human approval for high-risk actions like eval or download.\n- [DATA_EXFILTRATION]: The skill handles authentication state and file downloads. It provides security best practices including the use of encryption keys for state files and ensuring sensitive data is not committed to version control.\n- [PROMPT_INJECTION]: As the skill interacts with untrusted external web content, it is subject to Indirect Prompt Injection. Documentation provides mitigation through the AGENT_BROWSER_CONTENT_BOUNDARIES and AGENT_BROWSER_ACTION_POLICY features.\n
Ingestion points: Web content is processed via commands like snapshot, get text, and eval.\n
Boundary markers: The skill supports a --content-boundaries flag to delimit external content for the LLM.\n
Capability inventory: The tool can perform actions such as eval, download, and network routing.\n
Sanitization: Governance is implemented through domain allowlisting and an action policy framework.\n- [EXTERNAL_DOWNLOADS]: Dependency installation instructions target the official NPM registry and trusted repositories on GitHub.

run-agent-browser