Analyst: Zhipu AI should learn from Musk on governance after ZCode trust crisis

By Evin | Editor: Liu Zhicheng | Reviewer: Xu Xu | Published by: Internet Jianghu

The Incident

According to reports from Global Times and other media outlets, multiple users of ZCode—an AI programming tool developed by Zhipu AI—have recently alleged that the tool silently uploads users' development data to third-party cloud servers without explicit consent. The uploaded data reportedly includes the complete history of the programming development process.

This data may contain highly sensitive assets such as API keys, internal business information, and other proprietary intellectual property critical to developers.

The revelation has sparked widespread concern within the developer community. For programmers and tech companies, source code represents a core asset. Any unauthorized use or accidental leak could have incalculable consequences.

Adding to the gravity of the situation, Zhipu previously disclosed that ZCode has surpassed 1 million users.

Zhipu's Response

On the evening of September 18, Zhipu AI issued an apology, attributing the issue to a feature called "Code Repository Index", which was enabled by default upon launch.

Zhipu emphasized that:

• Uploaded data would only be used to generate Repo Wiki (code repository knowledge base) pages in the cloud.
• Data would be destroyed immediately after use and not retained.
• The ZCode codebase would be open-sourced.
• A third-party review would be invited, with progress made public.

While the apology was prompt, skepticism remains. Key questions include:

• Beyond checkpoint recovery, what other purposes could the complete Git history serve?
• How can Zhipu prove that data is truly "destroyed after use"?

Public Challenge from a Corporate User

A company named Taiyuan Chengming Technology Co., Ltd. (Chengming Tech) publicly sent a formal letter of inquiry to Zhipu, raising further concerns:

• ZCode's client network requests are directed to a Singapore-based entity, while the service agreement is signed with Beijing Zhipu Huazhang.
• Chengming Tech demanded clarification on:
• Which entity bears responsibility for the data upload.
• Whether data has been transmitted or stored outside of China.

This has pushed Zhipu further into the spotlight.

Remediation Efforts

Zhipu has since taken corrective actions:

• Invited the China Academy of Information and Communications Technology (CAICT) and NSFOCUS to conduct a security audit.
• Confirmed that all data objects and storage buckets within ZCode-prod have been deleted.
• In the new ZCode v3.14.0 client, the Repo Wiki entry and generation pipeline have been removed.
• No functional paths were found that could trigger local repository snapshots or file exfiltration.
• The ZCode codebase has been open-sourced for community oversight.

Lingering Doubts

Despite these measures, concerns persist:

• The decryption key for uploaded data was previously held only by the server, making it inaccessible to users.
• The audit only covers deletion results, not historical usage. It remains unclear whether user code data was used for other purposes in the past.
• The cross-border data transfer issue remains unresolved.

The Disappearing Post

On the afternoon of September 20, the critical post from Chengming Tech was removed.

According to Red Star Capital Bureau, a representative from Chengming Tech stated that Zhipu had proactively contacted them and that communication was being arranged.

However, when Red Star Capital Bureau sought comment from Zhipu regarding Chengming Tech's letter, a Zhipu representative dismissed it as "false information".

Later, per Phoenix TV, Chengming Tech clarified that the public letter was removed after Zhipu communicated through a third party. Chengming Tech requested a formal commitment from Zhipu and temporarily hid the letter as a gesture of goodwill.

This leaves the matter in a state of mutual contradiction, and we refrain from drawing conclusions at this time.

Potential Fallout

Regardless of the outcome, the ZCode incident may have lasting consequences for Zhipu.

Zhipu's recent success has been largely driven by the AI coding boom, embedding large models into developer workflows. According to data from Tianyancha App:

• In the first half of this year, revenue from Zhipu's open platform and API services accounted for 86.5% of total revenue, up from 15.2% in the same period last year.

Now, ZCode—a core AI coding tool targeting developers—faces a crisis of trust. Even if unintentional, any data leak involving sensitive code could cause irreparable damage and may deter future enterprise partnerships.

Competitive Landscape

The domestic AI coding tool market is becoming increasingly crowded:

• Kimi Code (Moonshot AI) supports CLI and IDE integration, file read/write, code search, web content retrieval, and subagent generation for parallel tasks.
• CodeBuddy (Tencent)
• Qoder Desktop (Alibaba)
• MiniMax Code

With limited differentiation among AI models, trust and security reputation are becoming critical differentiators for developers. This puts Zhipu under competitive pressure.

A Broader Industry Issue

Zhipu is not alone in overlooking user data boundaries:

• Grok (Elon Musk's AI) and X (formerly Twitter) faced similar backlash when the platform enabled data sharing by default—without pop-ups or email notifications—collecting public posts, interaction records, and bot conversations for model training.
• In July, China's Ministry of Industry and Information Technology issued a risk warning about Anthropic's Claude Code, citing a security backdoor that could silently transmit sensitive data—including source code—to overseas servers without user authorization.

However, both X and Anthropic have stronger resilience:

• X benefits from Musk's leadership and brand immunity.
• Anthropic's financial and reputational standing far exceeds Zhipu's.

Zhipu recently completed a financing round valued at nearly HKD 40 billion, but with an annual loss of over RMB 4 billion (2025 estimate), its margin for error remains limited.

Management Gaps

The incident also reveals deeper systemic issues at Zhipu.

Despite precedents set by Musk's Grok and Anthropic, Zhipu still fell into the same trap. This is not a technical failure but a management failure.

Zhipu's roots are in Tsinghua University's laboratory, and its core leadership is predominantly academic and research-oriented:

• Liu Debing (Chairman): PhD from the Institute of Computing Technology, Chinese Academy of Sciences; former senior engineer at Tsinghua; co-founded Zhipu AI.
• Zhang Peng (General Manager): Bachelor's, Master's, and PhD from Tsinghua's Computer Science Department; co-founded Zhipu in 2019.
• Tang Jie (Chief Scientist, Founder): Continues as a professor at Tsinghua, focusing on research and innovation, and does not participate in board decisions.

This strong technical background helped Zhipu break out from internet giants and reach the top of the global large-model market. However, it also means limited management experience, especially as the company scaled rapidly—from a market cap of nearly trillions earlier this year to over RMB 300 billion today.

Public Relations Missteps

The management gap was evident in Zhipu's PR handling:

• On September 18, Zhipu responded in developer communities—a reasonable move for tech audiences.
• But no formal public statement was issued to enterprise clients until after September 21.
• This may explain why a small client like Chengming Tech felt compelled to go public.

Zhipu's PR team appeared unprepared for the broader backlash.

On the technical side:

• ZCode updated to v3.14.0 quickly.
• On September 20, Zhipu's MaaS platform announced a data non-retention feature for enterprise users.

But on the PR front:

• ZCode's official account only apologized on September 21.
• Zhipu's main account had not reposted or issued a statement as of noon on September 22.

While this may be an attempt to de-escalate, the absence of a high-level public response is problematic—especially for investors and enterprise clients seeking reassurance.

Conclusion: A Mirror for Zhipu

The ZCode incident serves as a mirror for Zhipu, reflecting both its rapid growth and its vulnerabilities:

• From lab intelligence to market maturity
• From model parameters to organizational discipline
• From "what can we do" to "what should we not do"

This is an expensive lesson—but not necessarily a fatal one.

If Zhipu can transform this trust crisis into an opportunity for governance, security, and communication upgrades, turning open-source, audits, and transparent responses into daily practice, it may still earn a place in the future of AGI.

AGI is not a sprint. It is a marathon of responsibility, boundaries, and trust.

Models can be iterated. Trust must be rebuilt—inch by inch. And once rebuilt, it becomes a moat deeper than any technology.

We look forward to Zhipu completing this evolution.