The front foot just apologized, and the back foot was held accountable again.
This is what Zhipu has been experiencing in the past two days.
On the evening of September 19, Taiyuan Chengming Technology Co., Ltd. sent a letter to Zhipu, asking how much corporate data its AI programming tool ZCode uploaded, where the data went, and whether it was actually deleted.
Just one day ago, Zhipu just admitted that some functions of ZCode may indeed trigger the upload of user code repositories, and publicly apologized for this.
However, Chengming Technology questioned that the matter was far more than just “uploading part of the code.”
It stated in the letter that after independent evidence collection, ZCode uploaded not scattered code fragments, but may be a complete project archive file, including complete source code, system architecture, Git version history, database passwords, interface keys, cloud service credentials and other sensitive information.
This also brought to the surface again several key issues that had not been clarified in Zhipu’s previous apology.
Picture/picture insect creativity
追責
The controversy first started with a 313MB file.
On September 18, technology blogger ferstar noticed that ZCode left an encrypted snapshot locally while cleaning up computer space. The file itself cannot be decrypted directly, but the file list left during generation shows that it contains 42,411 files, and the “.git” directory alone accounts for 86.6% of the total data volume.
Git can be simply understood as a set of code modification files. It not only records what the project looks like now, but may also save past versions, deleted files and modification traces.
For users, this means that there may be hidden features, internal system structures, and information such as keys, configurations, undisclosed features, etc. that have been removed from the current code.
Ferstar further analyzed that ZCode will make a snapshot of the workspace, encrypt it and upload it to the cloud; there is no obvious entrance in the user interface to turn off this mechanism. Since the public key used for relevant encryption is issued by the server and the private key is held on the server side, even if the user intercepts the locally generated encryption package, he cannot confirm what content it contains.
Since then, the controversy has quickly spread throughout the developer community.
That night, Zhipu apologized through the ZCode official group and explained that the problem originated from the code base indexing function. This function is used to generate a warehouse index locally to support functions such as session checkpoint recovery, historical version rollback, and Repo Wiki.
Among them, Repo Wiki can be simply understood as letting AI automatically generate a project description after reading the entire project. Zhipu acknowledges that the upload of warehouse data may be triggered when generating Wiki pages.
Zhipu also stated that because this function was enabled by default in the initial stage of launch, some users were affected, and the related problems have been fixed. It will also open source the ZCode code base in the near future and invite third-party evaluators to review the system operation.
But Chengming Technology was not convinced by Zhipu.
One day later, it sent a letter asking Zhipu to explain the specific uploaded content, and to disclose the whereabouts of relevant data, storage status, whether it is shared with third parties or used for model training, and to provide a data processing list, access records and deletion certificates. The company also reserves the right to make claims, complain and file lawsuits.
This means that the ZCode controversy is moving from the developer community’s questioning of product design to formal questioning of companies’ responsibility for data processing.
信任危機
This letter brought to the fore some vague points in Zhipu’s previous apology.
The first question is: where exactly are the upload boundaries?
Zhipu admits that when Repo Wiki generates a page, it may trigger the upload of warehouse data, but it has not yet explained what data needs to be read and uploaded to achieve this function.
Angel investor and artificial intelligence expert Guo Tao told China News Weekly that for AI programming tools, “reading the entire code base” and “uploading the code to the cloud” are two completely different security levels. When reading locally, the original code remains on the user’s device; once uploaded to the cloud, the code, configuration files and other data enter the manufacturer’s server, which further involves data collection, storage and circulation.
In other words, just because the user allows the AI to look at the code does not automatically mean allowing it to take the entire project out of the computer.
Guo Tao believes that for a more reasonable product design, priority should be given to completing code indexing locally, uploading only the necessary code context required for model inference, and excluding sensitive files such as Git history, keys, and environment configurations by default.
This is also one of the core issues of Chengming Technology’s accountability, that is, whether the range of data actually processed by ZCode exceeds the boundaries that users can usually expect.
The second question is: How to prove “immediate destruction”?
Zhipu said that after the Wiki page is generated in the cloud, the relevant uploaded data will be “destroyed immediately and will not be saved.”
However, once the data leaves the local area, it may go through multiple links such as original uploaded files, cloud decompression, intermediate caching, operation logs, and backups. Currently, Zhipu has not publicly stated what links are covered by the so-called immediate destruction, and whether companies can query historical upload records and obtain verifiable deletion certificates.
Guo Tao said that companies cannot judge whether data is actually deleted based on verbal promises from manufacturers. Deleting the database mark does not mean that the data is completely cleared. There may still be copies in the underlying storage, cache and backup. To make deletion truly verifiable, at least operation logs covering the entire process of uploading, processing, and deletion should be provided, and third parties should be allowed to conduct independent security audits.
This is why Chengming Technology requires Zhipu to provide data processing lists, access records and deletion certificates.
The deeper question is how companies will trust such tools in the future.
In the past, AI programming tools were more like chat assistants. Programmers posted code and AI gave answers. Today’s Coding Agents have begun to directly enter the development environment. In order to understand the project, modify the code, run commands and even automatically complete tasks, they need to obtain more and more local permissions.
As a result, enterprises are not only concerned about whether the model capabilities are strong enough, but also about safety.
Guo Tao said that whether the code can be processed locally, whether sensitive files are blocked by default, how data is retained, and whether permissions and key operations can be audited will become increasingly important thresholds for enterprises to deploy AI programming tools.
After all, these issues are directly related to the company’s R&D assets and trade secrets.
What ZCode’s turmoil exposed is an unavoidable threshold when AI programming moves from personal tools to enterprise production environments – the greater the authority, the more manufacturers need to prove that they only take the necessary data, and that every step is checkable, controllable and traceable.
For Zhipu, if we want to truly end this crisis of trust, we still need to make this issue clearer.
Reporter: Shi Hanxu
Related Reading
- Why is Xiaomi’s business in India becoming more and more difficult to calculate?2026-09-21
- Using AI to “resurrect” family members doesn’t seem to be a good business2026-09-21
- Mr. Zhihui’s love letter has been officially sent out2026-09-21
- Original: The long-term pain of Cyrus is worse than the short-term pain2026-09-21
- In order not to say goodbye to their lover, they began to study the code.2026-09-20
