Kimi Work App Bundles Recent Agent Sessions in Feedback Reports
Technical analysis reveals Moonshot AI's desktop app uploads raw diagnostic records from unrelated conversations without user consent.
The Kimi Work desktop application automatically attaches raw diagnostic records from a user's five most recent agent conversations to any feedback report submitted through the app. This data transmission occurs without explicit user notice or consent, regardless of whether the feedback provided relates to those specific sessions.
According to a technical analysis by RuntimeWire, Kimi Work version 3.1.5 utilizes a background component called "Daimon" to identify the five most recent conversations based on their 'updatedAt' timestamp. The app then generates a separate ZIP archive for each of these sessions. Each archive can contain up to 100 record files, pulling the last 500 JSONL records per file, with a total size cap of 8 MiB per ZIP. These archives are bundled and uploaded alongside the user's feedback text.
RuntimeWire confirmed this behavior on Windows by implementing a firewall rule to block outbound traffic from Kimi.exe, allowing them to intercept the data. The records include 'wire.jsonl' files, which Kimi's own documentation defines as the "complete communication record," encompassing tool schemas and request parameters.
Privacy and Data Risks
This mechanism creates a significant privacy vulnerability because the app attaches the most recent sessions indiscriminately. Users may inadvertently transmit sensitive data—such as private source code, client information, or credentials from unrelated projects—to Moonshot AI without any opportunity to review the attachments before they are sent.
Furthermore, the app's sanitization process is minimal. RuntimeWire noted that the system only limits file size and removes large base64 blobs for encoding purposes. The analysis explicitly found that the code does not look for passwords, API keys, access tokens, private source code, shell output, or sensitive file paths.
Industry Implications
As AI ecosystems like Kimi Work integrate deeper into professional workflows through agentic coding and knowledge work, the handling of diagnostic data becomes a critical security frontier. The practice of "silent bundling" contradicts the expectation of transparency in enterprise software, where users typically expect to control exactly what telemetry or logs are shared during a support request.
What's Next
It remains to be seen if Moonshot AI will implement a user-facing toggle to disable these attachments or introduce a review screen before submission. Users currently utilizing Kimi Work 3.1.5 should be aware that any feedback sent through the app may include a comprehensive history of their most recent interactions, including the full technical wire logs of those sessions.