The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the codex CLI with the --dangerously-bypass-approvals-and-sandbox flag. This flag explicitly disables security isolation and bypasses human-in-the-loop approval prompts within the Codex environment, allowing for potentially dangerous actions to be performed without oversight.
[COMMAND_EXECUTION]: The execution path for the browser control script is determined by the BROWSER_SCRIPT environment variable. This allows an attacker who can influence the process environment to redirect execution to an arbitrary local script, leading to unintended command execution via Node.js.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. 1. Ingestion points: Screenshots of arbitrary web pages are captured in vision-click.mjs. 2. Boundary markers: No delimiters or 'ignore' instructions are present in the prompt sent to the vision AI. 3. Capability inventory: The skill has the capability to perform mouse clicks on the user's screen via browser.mjs. 4. Sanitization: There is no validation or sanitization of the screenshot content or the user-provided target string.
[PROMPT_INJECTION]: User-provided descriptions for the target element are directly interpolated into the prompt sent to the vision AI. This allows for direct prompt injection where a malicious description could override the AI's intended logic.
[EXTERNAL_DOWNLOADS]: The skill documentation requires the manual installation of playwright-core, which is a well-known and trusted library for browser automation.

vision-click