The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The file webbridge-setup.md instructs the agent to download and execute shell scripts directly from a remote server using the dangerous curl | bash and irm | iex (PowerShell) patterns. This occurs during the setup validation phase (Step 6) to install a browser automation tool. Evidence includes curl -fsSL https://cdn.kimi.com/webbridge/install.sh | bash and irm https://cdn.kimi.com/webbridge/install.ps1 | iex.
[EXTERNAL_DOWNLOADS]: The skill downloads and executes scripts and binaries from cdn.kimi.com, which is not an organization listed in the trusted vendors list. These downloads are performed without integrity verification such as hash checking.
[COMMAND_EXECUTION]: The skill executes a broad range of system commands and third-party CLI tools, including git, gh (GitHub CLI), glab (GitLab CLI), and project-specific build/test scripts detected in the repository such as npm, cargo, and vitest.
[PERSISTENCE]: The skill installs a background daemon (kimi-webbridge) on the local machine and ensures it is running using the start command. This daemon is designed to remain active across sessions to provide the webbridge functionality for browser-based tasks.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it processes untrusted data from the repository's own files to influence its behavior and configuration suggestions. Ingestion points include manifest and configuration files like package.json, Cargo.toml, and CODEOWNERS. No boundary markers or sanitization are applied to the ingested strings before they are presented to the agent or written to configuration files. The skill has extensive capabilities including network access and command execution which could be triggered by injected content.

configure-repo