Economic Observer reporter Ren Xiaoning

On the afternoon of September 21st, the stock price of Zhipu (02513. HK) fell by 5%. On the same day, Zhipu announced that ZCode (AI code editor under Zhipu) had completed the rectification, promising that the relevant code data was not used for model training, apologizing again, and open sourcing ZCode. In the afternoon, the stock price of Zhipu began to rise.

On the evening of September 18th, ZCode was exposed for uploading user code data to cloud servers without the user's permission. This incident has sparked controversy among the developer community, who believe there is a risk of information leakage.

It's like you paid for a cleaner, but without your knowledge, the cleaner searched your house and took away the copies of your property certificate and bank card password that you had put in the safe. "This is how the chief scientist of an AI company described the act of uploading user data.

In the following three days, Zhipu took multiple rounds of remedial actions.

In the field of AI coding, similar behavior has occurred twice this year. In July of this year, Musk's company xAI's Grok Build was exposed for packaging user data and uploading it to Google Cloud. Eventually, Musk publicly admitted and promised to delete the data, and open sourced Grok Build. Open source means that products may be more easily subject to user supervision.

A domestic big model practitioner told reporters that Meta has a plan that if users allow Meta to use their data, Meta can provide a particularly cheap package.

On September 21st, Zhipu announced the open source release of ZCode. However, some developers have found that ZCode is not fully open and still retains a closed source version. In addition, the open-source version of ZCode cannot enjoy discounts and token benefits.

At present, programming is one of the mature application scenarios for AI big models for enterprises, with a large number of willing to pay enterprise users, and it is also a battleground for big model companies. But enterprise customers have much higher requirements for security than consumer customers.

A company with more than ten employees used to have 80% of its code generated by AI, mainly using the Zhipu big model product. Now, the company has banned all employees from using ZCode.

Zhipu is one of the leading AI modeling companies in China, with a market value of HKD 1.33 trillion at one point.

The beginning and end of the storm

On the evening of September 18th, the chief scientist mentioned above paid attention to the Zhipu ZCode incident. ZCode is his main product for coding, and he immediately checked whether his data had been uploaded.

That evening, tech blogger Ferstar posted that as long as the user logs into their ZCode account, the software will package and encrypt the entire project workspace and complete Git folder in the background, and then transmit them to the cloud. The packaged files are "almost entirely core assets".

The Git folder is a complete version history repository for a project, containing commit records, branches, tags, configurations, and more. Its "information density" is extremely high, and it may contain sensitive information such as keys, passwords, internal addresses, etc. The chief scientist mentioned above gave an example that there may be a Bitcoin key in a Git folder, and if it is leaked, it could mean that the user's property may be damaged.

After investigation, the chief scientist mentioned above found that his data was not leaked because he did not log in to his account every time he used ZCode and only used it locally. But on social media platforms, there are constantly developers posting about their data being uploaded by ZCode, and some developers have found that even if the data upload switch is manually turned off in the ZCode settings interface, the background process still uploads data. A developer told reporters that he immediately used an agent to verify after the incident and found that there was indeed data uploading in the background snapshot, and the process was packaged in plaintext.

On the evening of the 18th, Zhipu ZCode officially responded and apologized, admitting that the problem was caused by the default enabled "code base index" function, and stating that the uploaded data will only be used to generate Repo Wiki (code repository knowledge base) pages in the cloud, and will then be destroyed and not saved. Zhipu also stated that the relevant issues have been fixed.

The developer community has raised three questions. One is to question the motivation behind Zhipu uploading user data; Secondly, some users discovered that even after Zhipu apologized, their upload behavior could still be monitored in the early hours of the same day; Thirdly, the claim that Zhipu destroys data lacks third-party authentication, making it impossible for users to confirm its authenticity.

Some developers have questioned whether Zhipu uses user data to train models. "Using the user privacy data training model is a very cheap way to obtain high-quality data." A technician of a large model of an Internet company said that this behavior is difficult for users to obtain evidence, and it is difficult to directly restrict it through laws and regulations.

On September 21st, Zhipu made a clear response to the query, stating that "Zhipu promised not to retain the code data mentioned by the community, and has never used it for model training

On the same day, Zhipu apologized again and made further responses, announcing that it had completed the rectification and open sourced ZCode to GitHub for community supervision. It also invited the China Academy of Information and Communications Technology and Green Alliance Technology to conduct security audits.

The crisis of trust is spreading

Zhang Guo, the founder of Liangchao Technology, is a user of Zhipu GLM5.3. He has a company of more than ten people, and 80% of the code is generated by AI. Now Zhang Guo has banned all employees from using ZCode.

In August of this year, Zhang Guo migrated the company's main model from DeepSeek to Zhipu. At that time, he recommended Zhipu to the Economic Observer, believing that Zhipu's model had stronger performance and higher cost-effectiveness.

A responsible manufacturer should minimize the storage of sensitive customer data in their own system, "said Zhang Guo. His company mainly provides enterprise services and pays high attention to customer data security. He has previously experienced incidents of cloud service providers leaking and shifting blame, and has not overestimated the social responsibility of large model enterprises.

A user posted on social media that after the incident, the company banned ZCode and even stopped using the cloud service provider that used ZCode to upload data.

Although the data of the chief scientist mentioned above has not been uploaded by ZCode, he has become more cautious and is using a MacBook for local deployment. When it comes to personal privacy data such as medical examination reports, he will use model analysis on this MacBook. In the future, he will also do more localized deployments to reduce the usage of cloud based large models.

Some companies have already taken precautions against privacy and security issues related to large models.

The CTO of a listed company in Guangdong told Economic Observer that he has strict regulatory requirements for the use of AI products within the company, and will clearly inform employees which jobs are suitable for using models, how to use them, and what permissions can be granted.

He asked employees not to use AI products for the entire project, but to use models for coding in fragments or in some projects, and not to give all their data to the model.

In addition, he will also require employees to enable privacy mode and not share data with the model. In terms of specific configuration, the entire project cannot be placed within the readable range of the model, only allowing the model to read the current directory. Git folder directories, key configurations, sensitive folders, etc. must not be read by the model.

At the end of the day, we need to start with development management standards. Giving models to all engineering projects is equivalent to having no management

Influence

After the incident, Zhipu took multiple measures: apologizing twice, resetting the credit limit for all ZCode users, inviting the China Academy of Information and Communications Technology and Green Alliance to conduct security audits, open-source ZCode, announcing the MaaS platform's "data content not retained" function, and establishing a vulnerability feedback and reward mechanism.

The current market value of Zhipu is HKD 387.1 billion, which has decreased by nearly HKD 1 trillion compared to its peak of HKD 1.33 trillion. After the outbreak of the user data upload incident, some companies have banned ZCode, which may have a revenue impact on Zhipu.

ZCode to Zhipu is similar to Chrome to Google. ”The chief scientist mentioned above said. ZCode is a product under Zhipu and will list Zhipu's model as the default model. If users use other coding products, the probability of the Zhipu model being recommended for use will decrease, which may affect Zhipu's API revenue.

API revenue is currently one of the main sources of income for Zhipu. According to the 2026 interim report of Zhipu, its total revenue for the first half of the year was 954 million yuan, of which revenue from open platforms and API services was 825 million yuan, accounting for 86.5%.

Other competitors are also taking action. On the evening of September 18th, MiniMax announced the open source of its AI coding product. There are currently over 20 mainstream AI coding products, which is also the largest revenue and most competitive field in the AI industry.

On September 21st, Dongwu Securities released a research report, maintaining the "buy" rating of Zhipu. The bank pointed out that Zhipu has completed the rectification of ZCode product security issues and made ZCode open source. It is expected that this incident will have a minor impact on the fundamentals.

Dongwu Securities also mentioned in its research report that in mid September, the number of ZCode users doubled from 1 million in August, and it is currently one of the top three coding agents in China.

(Our reporter Liu Sixuan also contributed to this article)

(Author Ren Xiaoning)