Zhipu open-sources ZCode — but only the version after the fix

Zhipu did the thing it promised and published its coding agent's source. What it published starts after the bug.

Zhipu (2513.HK) pushed the full source of ZCode to a public repository on Sunday night, nine days after a developer showed the agent was packing entire workspaces and Git history into an encrypted archive only Zhipu could open. The code lives under the company's zai-org organisation, carries an Apache 2.0 licence, and declares version 3.14.0 in its package manifest. It drew roughly 2,100 stars and 520 forks within hours of going up, and the README walks through building the desktop app, the browser client and the terminal agent from source.

The history is the part that decides what the release is worth. The repository has two commits: an initial commit at 12:06 UTC and a single squash labelled "feat: open source" at 21:14 UTC the same day. There is no pre-fix code in it — no workspace snapshot uploader, no Repo Wiki generation pipeline, nothing that shows the behaviour ferstar reconstructed on his own disk. A scan of the published tree turns up no repository-snapshot or workspace-upload endpoint at all; the only upload-credential call anywhere in the codebase belongs to the feedback form's attachment path, which is a different mechanism entirely. Anyone who wanted to audit the leak cannot audit the leak, and the community said so within hours — the top reply on the announcement thread is blunt: publishing the fixed code is not open source, it is a press release with a licence file.

The third-party review that landed alongside it covers the same window. CAICT, the state research institute Zhipu invited in, tested the production object-storage bucket behind ZCode and reported it holding zero data, and the security firm NSFOCUS confirmed that the stored data objects and the bucket itself had been deleted. Zhipu's own account of the fix is narrower: version 3.14.0 removed the Repo Wiki feature and cut the local snapshot generation and upload path, and the MaaS platform will add a no-retention option for submitted content. We traced this incident when it broke — ZCode quietly uploaded whole repos and git history to Z.ai's cloud.

That is the shape accountability takes when the evidence is already gone. Every claim now on the record is present tense: the bucket is empty, the client no longer packs the workspace, retention will be off. None of it can be checked against the version that did the packing, because that version was never published — and the release note that mentioned optimising memory for repository snapshot uploads on September 16 has since been removed from the changelog. Releasing a fixed client proves the fix exists. It does not prove the exposure was bounded, and it does not answer who held the decryption key for the archives that already left users' machines. Open source is the right instinct pointed at the wrong artifact.

What to watch: whether the pre-fix history or a server-side audit of the upload path is ever published, and whether the next Chinese coding agent to ship its source includes the version that failed.

Does publishing the fixed code settle anything for you, or does accountability require the version that failed? Tell us in the comments.

Sources: zai-org/ZCode (GitHub)

Deep Dive — America's top open model still trails China's best

Meta released the weights for Muse Spark Open on Thursday and called it "the most capable US open-weight model to date." Independent measurement agrees with the first half of that sentence and complicates the second: on Artificial Analysis's Intelligence Index the model scores 39, first among American open weights and behind GLM-5.3 at 45, Kimi K3 at 44 and Alibaba's Qwen3.8 at 40. The qualifier in Meta's headline is doing the work. This is a real release, a real engineering achievement, and not

AI 101 — How do AI video generators work?

An AI video generator is a model that starts with a grid of random noise and repeatedly refines it into a moving picture that matches your prompt — the same trick image generators use, stretched across time. Nothing is filmed, and nothing is stitched together from clips the model has seen. The pixels in the finished video were invented, frame by frame, all at once. Why it matters right now Kling 4.0, Kuaishou's flagship video model, went fully live on October 15 with up to 30 seconds of vid

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Read more

Deep Dive — America's top open model still trails China's best

AI 101 — How do AI video generators work?

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR