Abstract:
Where is the user's data packaged and uploaded by Zhipu? Just after Zhipu apologized, the controversy over ZCode's "stolen data transfer" turned to "data export".
ZCode interface
On the evening of September 19, Taiyuan Chengming Technology Co., Ltd. (hereinafter referred to as "Chengming Technology") officially disclosed a letter sent to Beijing Zhipu Huazhang Technology Co., Ltd. (hereinafter referred to as "Zhipu"), disclosing more information on the code base data upload issue involved in ZCode, Zhipu's AI programming tool, and stated that it reserves the right to take further legal action.
The core assets of the entire project were almost stolen?
It is understood that the incident can be traced back to September 18. In the early morning of that day, ferstar, an individual developer, posted on the technology forum that he found a 313MB encrypted compressed package locally. The status file showed that the file included almost all the core assets of a 10G commercial project. However, due to upload failure, it had been stuck locally and had failed 564 times until it was discovered.
After reversing the ZCode client, ferstar found that when the user is logged in, the client will generate a workspace snapshot, including the complete Git version history, LFS cache, reflog and other contents, which will be encrypted locally and uploaded to Alibaba Cloud OSS by Zhipu.
The developer also mentioned that the RSA public key used in the encryption process is issued by the server. The locally generated encryption package cannot be unpacked by the user or the client itself. The upload mechanism is always on by default, and there is no switch in the client interface that can actually turn it off.
ZCode upload link restored by ferstar
After the article was published, it quickly attracted attention. Several users successfully reproduced it according to the ferstar method, confirming the existence of the mechanism.
In the afternoon of that day, Zhipu apologized through the ZCode user group. Zhipu explained that the problem stems from ZCode's "code base index" function, which is used to support session checkpoint recovery, historical version rollback and Repo Wiki (warehouse encyclopedia) and other functions. Repo Wiki may trigger the upload of warehouse data when generating pages.
Zhipu stated that because this function was enabled by default in the initial stage of launch, some users were affected, so we are deeply sorry
. However, Zhipu promised that after the Wiki page is generated in the cloud, the relevant uploaded data will be "destroyed immediately and will not be saved", and said that the relevant problem has been fixed.Zhipu also announced that it will open source the ZCode code base in the near future, introduce third-party evaluators to review the system operation, and provide all users with an additional weekly quota reset.
ZCdoe apologized within the user group. Picture from IT Home
Data is suspected of flowing to Singapore
However, Chengming Technology has questioned the actual upload range, flow direction and so-called repair time of the smart spectrum data.
Chengming Technology stated in the letter that through its own technical evidence collection, it was discovered that from August 28 to September 14, more than 6 workspaces of the company were uploaded to the cloud by ZCode, with the largest one reaching 391.94 MB in size.
In addition, the workspace of another project of Chengming Technology is currently packaged, but due to 32 failed uploads it has not yet been successfully delivered to the Zhipu Cloud, so it remains locally and has not been deleted. "This situation is sufficient to prove that your company's processing behavior is automatically triggered and occurs in batches, and is not premised on any express or implicit consent of the company."
Chengming Technology Douyin Account
Chengming Technology pointed out that the uploaded content was not a "code snippet" as Zhipu officially called it, but a complete archive file, including the project's complete source code and system architecture design, complete version control history (including deleted historical submission and reference logs), database passwords, interface keys, cloud service credentials and certificates, personal information of employees and end users, as well as undisclosed R&D plans and business plans, which far exceed the collection scope stated in the "ZCode Privacy Policy".
At the same time, Chengming Technology also disclosed an issue with the actual flow of data.
Chengming Technology discovered that
all network requests from the ZCode client are directed to the zcode.z.ai and cdn-zcode.z.ai domain names
. According to the contracting party and personal data controller of the Z.AI platform, JINGSHENG HENGXING TECHNOLOGY PTE.LTD is registered in Singapore. The privacy policy also states that "services are usually provided from Singapore" and "your personal data is usually processed in Singapore."According to Zhipu’s prospectus on the Hong Kong Stock Exchange, JINGSHENG HENGXING is an indirect wholly-owned subsidiary of Zhipu registered in Singapore in November 2023. In the English version of ZCode, it is clear in its agreement and privacy policy that JINGSHENG HENGXING provides services and is responsible for data processing. However, the Chinese version of the ZCode User Agreement stipulates that the service provider is "Beijing Zhipu Huazhang Technology Co., Ltd." and the privacy policy also stipulates that the data processor is "Beijing Zhipu Huazhang".
The Chinese ZCode privacy agreement also mentions that all personal information collected and generated by China in its operations within China is stored within China and will not be transferred or stored across the border. If cross-border transmission or storage is required, the user will be notified in advance and the user's consent will be obtained, except in cases where notification is not required by law.
ZCode Privacy Agreement
Therefore, Chengming Technology requires Zhipu to explain the actual data processing subject of ZCode and whether there is cross-border transmission or overseas storage.
In addition, in the public letter, Chengming Technology also made 11 major requests to Zhipu, including stopping relevant data processing, deleting uploaded data and cache backups, disclosing data processing and access records, explaining whether it is shared with third parties or used for model training, and providing deletion certificates, etc., and requires Zhipu to respond in writing before October 10.
Currently, Zhipu has not made any further response to Chengming Technology’s letter. However, ZCode has released version 3.14.0 on September 19, and the update log clearly mentions that "the problem of abnormal upload of the warehouse encyclopedia has been fixed."
Comments