The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The README.md and index.html files instruct users to install the skill using curl -fsSL https://raw.githubusercontent.com/BowTiedSwan/agent-browser-skill/main/install.sh | bash. This pattern is high-risk as it downloads and executes a shell script from a remote repository without prior verification.
[EXTERNAL_DOWNLOADS]: The install.sh script performs a global installation of the agent-browser package from npm. While the tool is attributed to a well-known organization, the installation script itself resides on a third-party repository.
[COMMAND_EXECUTION]: The skill is designed to run the agent-browser CLI tool, which allows the agent to execute subprocesses for browser navigation, interaction with elements, and file system writes for screenshots.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its core functionality of reading web data. Ingestion points: External web content enters the agent context through agent-browser open and snapshot as described in SKILL.md. Boundary markers: There are no boundary markers or explicit instructions telling the agent to treat retrieved web content as untrusted. Capability inventory: The agent can execute commands, write files (screenshots), and access the network. Sanitization: No sanitization, filtering, or validation is performed on the data extracted from the headless browser before it is provided to the agent for processing.

agent-browser