The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to have the agent read and act upon data from external, untrusted websites (e.g., snapshots and accessibility trees). This creates a risk where malicious instructions hidden in web pages could influence the agent's actions (e.g., via click or fill commands). There are no specified sanitization methods or boundary markers in the provided templates to mitigate instructions embedded in web content. (Found in SKILL.md, templates/ai-agent-workflow.md, examples/agent-mode/optimal-workflow.md).\n- [DYNAMIC_EXECUTION]: The eval command allows for the execution of arbitrary JavaScript within the browser context. This is a powerful capability that could be misused if the agent is manipulated into executing code sourced from a malicious webpage. (Found in api/commands.md, examples/commands/advanced-commands.md).\n- [COMMAND_EXECUTION]: The skill provides integration templates for shell-based automation (e.g., Python's subprocess.run(shell=True)). Users should ensure that any parameters passed to these shell calls are sanitized to prevent command injection, especially when incorporating data retrieved from the web. (Found in templates/ai-agent-workflow.md, examples/agent-mode/integration.md).\n- [PRIVILEGE_ESCALATION]: Documentation for system setup on Linux platforms includes sudo apt-get install commands for installing browser dependencies. Users should verify these requirements before execution in their environment. (Found in examples/getting-started/installation.md).\n- [EXTERNAL_DOWNLOADS]: The skill includes examples of downloading resources from the Vercel Labs GitHub organization using curl. These references target an official repository associated with the tool's developers and are categorized as safe references to a well-known service. (Found in examples/advanced/usage-with-agents.md).

agent-browser