The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The skill relies on the agent-browser CLI. The provided metadata references (https://github.com/AnsarUllahAnasZ360/cc-guide and https://agent-browser.dev) are not associated with any Trusted Organizations. Use of tools from unverified sources increases the risk of supply chain attacks.
COMMAND_EXECUTION (MEDIUM): The agent-browser eval command allows the agent to execute arbitrary JavaScript within the browser context. While a standard feature for automation, it presents a risk if the agent is manipulated into executing code that targets the user's active sessions or local environment.
CREDENTIALS_UNSAFE (MEDIUM): The documentation explicitly demonstrates retrieving sensitive data via agent-browser eval "localStorage.getItem('token')". This pattern encourages the extraction of session credentials, which could be easily exfiltrated if combined with network tools.
PROMPT_INJECTION (LOW): The skill is highly vulnerable to indirect prompt injection from malicious web content.
Ingestion points: agent-browser open, agent-browser snapshot, and agent-browser get text allow untrusted data from external websites into the agent's context.
Boundary markers: Absent. There are no instructions or delimiters to help the agent distinguish between its own instructions and the content retrieved from the web.
Capability inventory: The skill possesses Bash tool access, browser eval capabilities, and network routing control (agent-browser network route).
Sanitization: Absent. The skill does not perform any validation or filtering on the data scraped from web pages before the agent processes it.

agent-browser