The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the belt CLI tool to perform browser automation, which involves executing shell commands to interact with the remote inference.sh browser service.
[REMOTE_CODE_EXECUTION]: The execute function allows the agent to run arbitrary JavaScript code within the browser session, providing a powerful vector for data extraction or DOM manipulation that could be exploited if triggered by malicious inputs.
[CREDENTIALS_UNSAFE]: The skill supports automated authentication flows. Although the documentation recommends using environment variables and secret managers, the handling of credentials by an agentic tool involves inherent risks of exposure.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes content from external websites. Attackers could embed instructions in web pages to manipulate the agent's behavior. 1. Ingestion points: The snapshot and execute functions, and extraction templates like capture-workflow.sh. 2. Boundary markers: None. 3. Capability inventory: belt CLI execution, file upload actions, and arbitrary JavaScript execution. 4. Sanitization: No explicit filtering or sanitization of external web content is performed.
[EXTERNAL_DOWNLOADS]: The skill references a CLI installation script from the inference-sh vendor's GitHub repository, which is a required component for the skill's operations.

agent-browser