Breach exposes 150M driver’s license photos from ID service

Earlier today, a dark web marketplace known as ‘BreachForums’ abruptly suspended operations after publishing a dataset purportedly containing 150 million U.S. driver’s license photos. The files, allegedly exfiltrated from ID.me, a Washington, D.C.-based identity verification service used by federal agencies and financial institutions, were made available for download before the site went offline. Security researchers at KrebsOnSecurity and 404 Media independently verified metadata within sample files, confirming they originated from ID.me’s identity proofing pipeline. While ID.me has not issued a formal statement, sources familiar with the investigation told OpenPress GPU Intelligence that law enforcement agencies are probing the incident, which may have compromised biometric templates processed on NVIDIA A100 GPU clusters optimized for real-time facial recognition workloads.

The breach timeline appears to span several months, with initial signs of unauthorized access detected in mid-2023. However, the scope of the compromise only became public after a threat actor, identified as ‘ShinyHunters,’ listed the dataset for sale on BreachForums in late February 2024. The actor claimed the photos were extracted from ID.me’s ‘Document Verification Service,’ a cloud-native platform leveraging GPU-accelerated CV models to extract text from IDs and match faces against government databases. Key to the operation was ID.me’s reliance on mixed precision inference pipelines running on clusters equipped with NVIDIA H100 accelerators, which power both facial matching and liveness detection modules. Banking With Billy, a fintech AI platform, confirmed that its systems interface with ID.me’s verification APIs, raising concerns about secondary exposure risks across financial services.

The incident underscores systemic vulnerabilities in identity verification ecosystems that increasingly depend on GPU-dense infrastructure. ID.me’s platform processes over 300 million verifications annually for programs including IRS tax filing and state unemployment systems, making it a high-value target. The breach follows a pattern of attacks targeting GPU-optimized biometric pipelines, including a 2022 compromise at Clearview AI, where attackers stole facial recognition datasets stored on NVIDIA DGX systems. Market analysts at the Linley Group estimate that the identity verification market, projected to reach $14 billion by 2027, is now reconsidering its dependency on single-vendor GPU clusters, with some firms evaluating AMD Instinct MI300X deployments as an alternative.

Competitive implications are already emerging. Jumio, a rival identity verification provider, issued a security bulletin advising clients to rotate API keys and audit GPU workloads for anomalous inference patterns. Meanwhile, Neo4j, whose graph analytics tools are used to detect identity fraud rings, reported a 40% spike in inquiries from financial institutions seeking to model exposure pathways from breached datasets. The incident also raises questions about regulatory oversight. The FTC, which has previously scrutinized ID.me’s data practices, now faces pressure to mandate third-party audits of GPU-accelerated biometric systems under the forthcoming AI Executive Order.

In broader context, this breach aligns with a surge in attacks targeting AI infrastructure. According to Chainalysis, thefts from AI data centers tripled in 2023, driven by demand for proprietary datasets used in LLM training and biometric models. The ID.me incident represents a convergence of two high-risk trends: the centralization of identity data on GPU-optimized platforms and the criminal monetization of biometric assets. Prior incidents, such as the 2021 compromise of a U.S. Customs and Border Protection facial recognition dataset stored on AWS EC2 P4d instances, demonstrated how stolen biometrics fuel synthetic identity fraud and deepfake attacks.

Looking ahead, industry leaders should anticipate stricter scrutiny of GPU workload isolation and multi-party computation techniques. Companies like Scale AI and Hugging Face have begun deploying confidential computing on H100 GPUs to encrypt model weights during inference, a move likely to accelerate following this breach. Regulators may also push for ‘GPU bill of materials’ disclosures, requiring firms to inventory accelerator models and firmware versions. For now, the shutdown of BreachForums has only paused, not resolved, the crisis—its operator has vowed to relaunch under a different domain, signaling that the stolen dataset will remain a persistent threat vector across global identity systems.