A San Francisco startup in human resources, Rippling, is causing concern due to its use of an AI model linked to a blacklisted Chinese lab. This comes at a time when U.S. companies face increased scrutiny from Beijing.
Rippling positions itself as a comprehensive AI solution for human resources, IT, and finance, serving over 20,000 companies in the U.S., according to its website. The model in question, developed by Z.ai, also known as Zhipu, had previously been placed on a U.S. government blacklist for its relationships with the Chinese military.
Ryan Fedusiak, a fellow at the American Enterprise Institute focused on China and AI, expressed serious concerns about this situation. He stated, “Any data that passes through servers in China is vulnerable to data theft and surveillance. This is akin to the issues we’ve encountered with TikTok but significantly escalated, as users are sharing sensitive personal data instead of just lighthearted videos.”
Fedusiak noted that open models like this can be manipulated in ways that are hard to detect, warning that while there haven’t been many attacks yet, the risk certainly exists.
Rippling is not alone in exploring Chinese AI capabilities. Companies ranging from DoorDash to various startups are increasingly leveraging these models, which offer nearly top-tier performance at significantly lower costs than competitors like Anthropic and OpenAI, as reported by Fortune. In 2025, Airbnb’s CEO described Alibaba’s Qwen AI as “fast and cheap.”
Rippling has not responded to inquiries about how extensively they utilize these Chinese AI models or the measures they take to protect their users’ data. Similarly, Zhipu has remained silent on the matter.
‘85% Cheaper’
In July, Parker Conrad, CEO of Rippling, praised Zhipu’s GLM 5.2 model, claiming it performed nearly as well as more expensive models but at a fraction of the cost—85% cheaper, to be exact.
Conrad tweeted, “This is why access to open weights matters – GLM 5.2 is 85% cheaper but nearly identical performance.”
According to an AI usage guide on Rippling’s website, the company mainly relies on OpenAI and Anthropic for its AI services and prohibits these providers from storing or utilizing client data for model training.
This is why access to open weights matters – GLM 5.2 is 85% cheaper but nearly identical performance. Really impressive to see how well Grok 4.5 perform here, on a set of benchmarks that labs are not training against.
— Parker Conrad (@parkerconrad) July 30, 2026
In the same month, Rippling President Matt MacInnis published results indicating that the company employed the contentious Z.ai GLM model during its evaluations. He mentioned that the benchmark tests were based on genuine business tasks involving real company data, revealing that the open-weight models were surprisingly effective across different measures, including cost efficiency.
Albert Strasheim, the company’s CTO, later conducted a “security review” using the GLM 5.2 model.
Furthermore, Rippling has been robustly engaged in lobbying efforts, securing tens of millions in tax incentives from various Democratic leaders while functioning as a significant vendor among Democratic political networks.
Some of Rippling’s client base includes companies operating in sensitive sectors, such as national security and defense.
For instance, Askari Defense, which develops anti-drone technologies and counts Rippling as a client, was selected as a semifinalist in a U.S. Army competition aimed at advancing defense technologies.
‘People Shouldn’t Be Blind To It’
In January, Zhipu was added to the Bureau of Industry and Security’s Entity List due to its potential to support China’s military advancements. This designation enables the U.S. to impose certain export restrictions on the company.
Firms are still allowed to use open Chinese models as long as sensitive information, like proprietary training data, isn’t transmitted to Zhipu.
Recently, Z.ai announced the release of its ZCode coding assistant following adjustments related to user data security concerns. ZCode operates on the updated GLM-5.3 model and has promised increased transparency.
The cybersecurity firm NSFOCUS noted that user code data had been deleted, as per Zhipu’s claims. This raises concerns given prior incidents associated with data mishandling.
A BIS spokesperson referred inquiries related to their annual report, highlighting that although a significant number of licenses were granted to companies on the Entity List, very few were issued in the latter part of 2025.
Additionally, Texas Governor Greg Abbott categorized Zhipu as a prohibited technology earlier this year to safeguard state systems from foreign threats.
Abbott remarked, “We should protect our networks from hostile entities that may misuse AI and other technologies.”
Reports suggest that Chinese AI models have surged to become a dominant force on major AI platforms, indicating their increasing capability and competitiveness in the global market.
This raises questions as to how businesses should approach partnerships involving Chinese technology. Chan advised that while connecting directly to these models poses significant risks, using intermediary platforms might reduce those risks—albeit not eliminate them entirely.
Fedusiak cautioned that smaller firms often opt for more budget-friendly AI solutions. However, he stressed the need for awareness when selecting partners in this space to ensure they uphold their promised data handling practices.
He added, “Z.ai’s incident highlights the necessity for thorough scrutiny to prevent such occurrences.”
Reports have indicated that leveraging Chinese AI models can lead to unintended consequences, including the misuse of user data, which requires companies to be extremely cautious in their integrations.
