The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install the required daemon by downloading a shell script from the vendor's delivery network at https://cdn.kimi.com/webbridge/install.sh and piping it directly into a shell interpreter.
[DATA_EXFILTRATION]:
The upload tool allows the agent to read files from the user's local filesystem and upload them to arbitrary external websites through the browser session.
The screenshot and save_as_pdf tools support a path parameter that allows the agent to write files to user-specified or computed local paths.
[COMMAND_EXECUTION]:
The skill executes and manages a local binary (~/.kimi-webbridge/bin/kimi-webbridge) to interact with the browser extension.
The evaluate tool provides the ability to execute arbitrary JavaScript code within the context of the active web page.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by reading external web content and providing tools to act on that content.
Ingestion points: The snapshot tool in SKILL.md retrieves the accessibility tree and page content from external websites.
Boundary markers: No specific delimiters or instructions to ignore embedded commands are present when processing page snapshots.
Capability inventory: The agent can ingest web data and subsequently execute actions via evaluate (JavaScript), upload (file reading), and navigate (network operations).
Sanitization: The instructions do not define sanitization or validation logic for data retrieved from web pages.

kimi-webbridge