The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends several commands that pipe remote shell scripts directly into bash for installation, such as curl -fsSL https://browser-use.com/cli/install.sh | bash. This practice executes unverified remote code in the user's shell environment.
[COMMAND_EXECUTION]: The browser-use python command allows the agent to execute arbitrary Python code strings. This provides a mechanism for running complex scripts that can interact with the local filesystem and network outside the intended browser automation scope.
[DATA_EXFILTRATION]: The skill facilitates the handling of sensitive browser session data through commands like browser-use cookies export and browser-use profile sync. The synchronization feature encourages uploading local browser profiles and authentication tokens to an external cloud service.
[PROMPT_INJECTION]: The skill is highly susceptible to indirect prompt injection due to its core function of reading and interacting with untrusted web content. Evidence chain: 1. Ingestion points: browser-use state and browser-use get html read external web pages into the agent's context. 2. Boundary markers: None identified to isolate web content from instructions. 3. Capability inventory: The agent can execute shell commands, run Python, and perform autonomous tasks via browser-use run. 4. Sanitization: No sanitization or filtering of webpage content is described.
[EXTERNAL_DOWNLOADS]: Fetches the cloudflared binary from Cloudflare's official GitHub repository to support local-to-cloud tunneling features.

browser-use