kimi-sync
Pass
Audited by Gen Agent Trust Hub on Oct 3, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes multiple local Python scripts using the
uvtool to perform administrative tasks. - Evidence includes calls to
scripts/kimi-audit.pyto compare local and remote skill lists andscripts/kimi-extract.pyto handle archive extraction. - It also performs
gitoperations, includinggit addandgit commit, to persist changes to the repository. - [EXTERNAL_DOWNLOADS]: The skill fetches data and files from external sources, specifically
kimi.com. - It downloads signed ZIP archives containing skill definitions and content using
curlor Python-based methods. - The instructions include a specific security check to reject archive entries that are absolute or contain path traversal characters (
..). - [DYNAMIC_EXECUTION]: The skill injects and executes JavaScript directly into the browser context to interact with the Kimi web application.
- It uses
browser_cdpwithRuntime.evaluateorbrowser_evaluateto run code fromscripts/kimi-list.jsandscripts/kimi-download.js. - This execution is used to scrape internal application state, such as the
VUE_QUERY_CLIENTcache and Vue provider objects, to retrieve skill metadata and download URLs. - [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection due to its core function of ingesting third-party skill data.
- Ingestion points: Data enters the environment from
kimi.comvia scraped skill lists, file tree metadata, and the content of downloaded ZIP files (SKILL.md). - Boundary markers: The skill implements path-based boundary checks (preventing ZIP slip) but lacks delimiters or warnings to the agent regarding the interpretation of natural language instructions found within the downloaded files.
- Capability inventory: The skill has capabilities including file system writes (
kimi-extract.py), browser manipulation, and repository modification (git commit). - Sanitization: While path validation is present for the extraction process, there is no evidence of content sanitization for the instructions embedded in the downloaded skill files.
Audit Metadata