The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The operations guide in references/operations.md includes instructions to install the bridge component using curl -fsSL https://cdn.kimi.com/webbridge/install.sh | bash.- [COMMAND_EXECUTION]: The skill executes a local binary ~/.kimi-webbridge/bin/kimi-webbridge for status and lifecycle management, and a shell script scripts/screenshot.sh to capture and process page images.- [DYNAMIC_EXECUTION]: The evaluate tool allows the agent to execute arbitrary JavaScript within the user's browser session, which can be used to interact with or extract data from complex web applications.- [INDIRECT_PROMPT_INJECTION]: The skill is inherently exposed to indirect prompt injection because it processes content from external websites. \n
Ingestion points: Tools like snapshot, evaluate, and network ingest untrusted data from the web. \n
Boundary markers: No specific delimiters are used to isolate ingested content. \n
Capability inventory: Extensive capabilities including navigation, form filling, and file creation. \n
Sanitization: External content is not sanitized before being processed.- [DATA_EXFILTRATION]: By design, the skill operates within the user's browser session and has access to 'actual login sessions.' Tools such as snapshot, network, and screenshot can expose sensitive personal or authentication data to the agent.

kimi-webbridge