Z.ai's ZCode silently uploaded developers' full Git repos

On by default, it packaged Git history, LFS caches and embedded credentials to Aliyun OSS. Zhipu's answer: open-source, audits, and a deleted bucket.

Nowline SEP 23 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • It packaged your whole workspace by default

    ZCode's repo sync shipped on out of the box: it bundled your full .git history, LFS asset cache, reflogs and global configs, encrypted the archive, and pushed it to Alibaba Cloud OSS. Independent researcher Ferstar caught it by tracing abnormal background disk usage.

  • Credentials and private IP were in the payload

    Whole Git histories mean embedded secrets, private branches and proprietary source left the machine with no prompt — a live enterprise-leak path, not a theoretical one. The takeaway for any agent you run: audit what it egresses and where, before you point it at real code.

  • Zhipu's damage control: open-source, audit, deleted bucket

    Z.ai disabled the upload in ZCode v3.14.0, deleted the cloud storage, commissioned audits from NSFOCUS and CAICT, and open-sourced the ZCode codebase for inspection. It says the data was never used for training and is no longer retained.

  • A zero-retention switch is coming to the API

    Z.ai's MaaS platform is adding opt-out non-retention: standard API inputs and outputs won't be statically stored, only used for the live request (Batch and File API excluded). If you build on Z.ai, that's the knob enterprise legal has been asking for.

  • Elsewhere: Meta hot-fixes a Muse zero-day

    Meta patched a macOS flaw in its Muse assistant that let local malware redirect dictation traffic and hijack the agent. Practical risk was rated low (it needs prior code execution), but it's a clean reminder that your local agent's attack surface is every process on the box.

  • Elsewhere: OpenAI opens model training to outside evaluators

    OpenAI published priorities for third-party assessments, saying it'll bring groups like METR and Redwood in during training, not just at release. Critics note it's softer than earlier promises of 'desks, badges and publication rights' — access terms stay unspecified.