As reported by SecurityAffairs, attackers compromised credentials belonging to Fireworks AI — a cloud platform for running and fine-tuning open AI models — and used that access to exfiltrate personal data belonging to approximately 149,346 Uber users, including driver's license images and government ID details. Uber severed the data connection on October 9, 2026, but the unauthorized access had occurred two days earlier on October 7. The breach was confined to Fireworks AI's environment and did not reach Uber's core systems.

Cloud Security Alert: As reported by SecurityAffairs, attackers compromised credentials belonging to Fireworks AI — a cloud platform for running and fine-tuning open AI models — and used that access to exfiltrate personal data belonging to approximately 149,346 Uber users, including driver's license images and government ID details.

Why This Matters Beyond Uber

This incident is not really about Uber's internal security — by all indications, Uber's segmentation held. The failure point is the AI infrastructure supply chain, and it exposes a structural problem that most third-party risk programs have not yet adapted to address.

Fireworks AI is an inference and fine-tuning platform. Companies send it data — sometimes highly sensitive data — to run models, generate embeddings, or fine-tune weights. In Uber's case, it appears driver's license images and identity documents were being processed through Fireworks AI's environment. That means Uber was effectively using a third-party AI vendor as a data processor for regulated PII, with the access profile of a cloud storage bucket but the security governance of an early-stage startup.

The attack surface is not the model. It is the credentials, the API keys, and the data residency controls around the model-serving infrastructure.

The Detection Lag Problem

Two days elapsed between initial unauthorized access (October 7) and provider notification (October 9). For a vendor holding images of government-issued IDs at scale, that is a meaningful window. It suggests Fireworks AI either lacked real-time anomaly detection on credential usage or did not have telemetry sufficient to identify unusual data access patterns quickly. Defender programs should treat this as a baseline expectation: if a vendor cannot detect and alert on anomalous API access within hours, it should not be handling identity documents.

The Pattern: AI Vendors as the New MOVEit

This breach fits a recurring pattern we have tracked across 2025 and 2026. Snowflake, MOVEit, and now Fireworks AI all share a common structure: a platform provider with broad downstream customer reach is compromised through credential theft or session hijacking, and the blast radius is determined not by the vendor's security but by how much sensitive data each customer deposited.

The difference with AI vendors is that customers often do not realize they are depositing sensitive data. Sending documents to a fine-tuning API feels different from uploading files to a file-sharing platform — but from a data exposure standpoint, it is functionally identical.

Shield53 Recommendations

What You Should Do

Shield53 Recommendations
Audit every AI vendor relationship for data classification. Identify which third-party AI or ML platforms currently receive PII, PHI, financial, or identity document data. If you cannot produce that list within 24 hours, you have a blind spot.
Enforce data minimization at the API layer. Before sending data to any AI inference or fine-tuning provider, apply redaction, tokenization, or synthetic substitution. Driver's license images should never reach a vendor environment in raw form if the use case does not strictly require it.
Require vendor detection SLAs in contracts. Contractual language should mandate notification of suspected credential compromise within 4 hours of detection — not 48. Include audit rights for access logging and anomaly detection controls.
Rotate and scope API credentials aggressively. Use short-lived tokens with least-privilege scopes. Long-lived API keys stored in vendor-side environments are the primary failure mode in these incidents.
Map AI vendor exposure to regulatory obligations. Under CCPA, GDPR, and U.S. state breach notification laws, the data controller (Uber in this case) bears notification responsibility even when the processor is the compromised party. Ensure your incident response plan accounts for processor-originated breaches.

The lesson here is not that AI platforms are uniquely vulnerable — it is that they are being trusted with sensitive data before the industry has matured its security practices to match that trust. Until vendor risk programs treat AI infrastructure providers with the same rigor applied to cloud storage and identity providers, these breaches will continue to scale with AI adoption.