The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's preflight check instructions direct the agent to surface a shell command that downloads and executes a script from an untrusted GitHub repository (voidmatcha/ui-clone-skills) using a curl | bash pipe. While the instructions explicitly state the agent should not auto-execute the command, recommending it to the user as the 'Fastest fix' facilitates high-risk remote code execution from an unverified source.
[COMMAND_EXECUTION]: The skill performs extensive command-line operations using agent-browser and ffmpeg to manipulate browser states and process video files. It also uses agent-browser eval to execute arbitrary JavaScript on third-party websites, which is a powerful capability that could be abused if the target site or the agent's logic is compromised.
[PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection because it ingests untrusted data from user-supplied URLs. Malicious actors could embed instructions within web page metadata, class names, or visible text that might be interpreted as commands by the agent when it processes browser evaluation results. Although the instructions mention redacting 'prompt-like text', this mitigation is incomplete and may not prevent sophisticated injection attacks.

ui-capture