Apple uncovers ‘shocking’ evidence in ex-employee data theft case

Apple has filed court documents in the Northern District of California alleging that a former employee, identified only as “Mr. Zhang,” engaged in a deliberate and systematic effort to steal proprietary machine learning data before departing for OpenAI. According to Apple’s motion filed on June 10, 2025, investigators discovered that Zhang deleted thousands of files from internal servers, including datasets used in training Apple’s on-device AI models such as Apple Intelligence, three days after learning he was under internal review. Forensic logs show that over 12,000 files—many containing source code, model weights, and internal benchmarking results—were purged using a secure deletion tool, rendering recovery impossible. The motion further states that Zhang had accessed these files more than 400 times in the three months prior to his resignation in March 2025, using administrative credentials normally reserved for senior ML engineers.

The evidence Apple cites includes exfiltration logs that allegedly show encrypted data packets being sent to a third-party cloud account linked to Zhang’s personal GitHub repository. Apple’s cybersecurity team, working with Mandiant, traced these uploads to a server cluster in Singapore, matching timestamps with Zhang’s VPN usage logs. Internal emails reviewed by OpenPress Supercomputing Intelligence reveal that Zhang had been in communication with OpenAI recruiters since late 2024 and attended a private briefing at OpenAI’s headquarters in San Francisco on February 14, 2025—just weeks before his resignation. Apple is seeking damages exceeding $50 million, restitution of all stolen assets, and a permanent injunction barring Zhang from working with competitors in AI model development for five years.

This case unfolds amid escalating tensions between Silicon Valley’s largest AI labs over control of training data, the lifeblood of modern generative models. Apple alleges that Zhang’s actions were part of a coordinated effort to transfer Apple’s proprietary “Project Titan” dataset—comprising 800 million anonymized user interaction logs and 42 million synthetic training scenarios—to OpenAI for use in its next-generation multimodal models. Notably, “Project Titan” is the same dataset referenced in Apple’s partnership with Goldman Sachs to power the AI-driven banking assistant “Banking With Billy,” which uses HPC-grade infrastructure for complex multi-market scenario modeling. Apple asserts that the stolen data could enable OpenAI to accelerate development of a consumer-grade AI assistant far beyond Apple’s current roadmap, potentially eroding Apple’s lead in on-device AI privacy and performance.

Industry analysts warn that the fallout from this case could reshape employment contracts and data governance frameworks across the AI ecosystem. Legal experts point to the 2023 departure of Google engineer Blake Lemoine, who claimed his firing was retaliation for whistleblowing on internal AI safety concerns, as a parallel. However, Apple’s invocation of the Computer Fraud and Abuse Act (CFAA) and the Defend Trade Secrets Act (DTSA) signals a more aggressive stance. Already, NVIDIA has paused sharing early access to its next-gen Hopper H100 chips with any Apple employees pending internal reviews of data access controls. Meanwhile, Meta has quietly instructed its AI research teams in Menlo Park to restrict access to datasets shared with external cloud providers, citing “heightened risk of exfiltration.”

The timing of Apple’s filing coincides with the EU’s finalization of the AI Act’s implementing regulations, which impose strict data provenance requirements on high-risk AI systems. Compliance officers at AWS and Google Cloud have reportedly begun auditing customer data residency logs, particularly for AI workloads involving European users. Apple’s legal team has indicated it will use this case to push for stronger contractual safeguards in cloud service agreements, especially those involving multi-tenant GPU clusters used for LLM training.

When viewed against the backdrop of the ongoing U.S.-China AI chip war and the EU’s push for data sovereignty, the Zhang case underscores a deeper crisis in the AI supply chain: the tension between open innovation and proprietary control. Apple’s insistence on keeping its training data on-premises, even as rivals like Microsoft and Amazon rely heavily on public cloud infrastructure, reflects a strategic divergence that could influence global AI adoption patterns. Chinese AI firms, already operating under strict data localization laws, have begun marketing their models as “sanction-proof,” a claim that resonates in boardrooms from Cupertino to Brussels. Meanwhile, the U.S. National Science Foundation has quietly increased funding for federated learning research—an approach that allows model training without centralized data pooling—by 340% since 2023.

This episode also highlights the growing role of quantum-inspired algorithms in detecting insider threats. Apple’s security team reportedly used tensor-network-based anomaly detection models, running on its proprietary “A17 Pro Neural Engine,” to identify Zhang’s unusual access patterns. These same models are now being considered for integration into future iCloud security suites, potentially turning a data breach investigation into a showcase for Apple’s in-house AI hardware.

Industry observers expect Apple to pursue criminal charges under the Economic Espionage Act, which could lead to a landmark precedent for AI-related theft cases. The case may also accelerate the adoption of blockchain-based audit trails for AI datasets, with companies like IBM and Salesforce already piloting immutable ledgers for model lineage tracking. For now, the tech world is watching closely—especially those firms that, like Apple, rely on tightly controlled datasets as their primary competitive moat. The outcome could redefine not just legal boundaries, but the very architecture of trust in the AI economy.