What happened
On 17 November 2026, Fireworks said its October activity involved two separate security incidents and broadened the confirmed data scope. The first involved unauthorised use of internal Fireworks credentials and access to environment variables used in evaluation jobs. The second began after a third-party cloud provider hosting Fireworks development machines experienced a security incident, allowing access to part of Fireworks’ production systems. Fireworks says both incidents are contained and remediated.
Fireworks says the second incident exposed account and billing information for all accounts created before 30 October 2026 and datasets, models and/or secrets for a small number of customers. It says payment-card numbers were not stored and were not affected. Fireworks also says it found no impact to Inference, Training SDK or Nexus, which it describes as zero-data-retention by default. Customer-specific impact still depends on direct notification from Fireworks.
Uber says the provider’s compromised credentials accessed its data on 7 October 2026; the provider notified Uber on 9 October, and Uber confirmed personal-information access on 13 October. Uber’s Iowa filing says the stolen dataset contained driver’s-licence images and text including name, address, date of birth and licence number. Reporting based on an Oregon filing put Uber’s total notified population at 149,346 people. Uber says valid provider credentials and permitted data flows meant its cloud alerting did not fire.
Fireworks says a third-party cloud provider suffered an incident on 23 October 2026, access reached Fireworks development machines on 24 October, and no further successful customer-platform access was detected after 25 October. Attribution posture: Fireworks links the activity to a financially motivated cybercriminal group associated with LLMJacking, but no named actor has been established. No public network indicator has been tied specifically to the Uber data access described in the regulatory notice. No AI agent or offensive framework was identified in the cited sources. No underlying model or model version was identified in the cited sources. The cited sources identify stolen credentials and LLMJacking as the objective, but do not publish an operator configuration. The observed mechanism was credential abuse against cloud storage and Fireworks systems; the cited sources do not establish that an AI system mechanically executed the intrusion. The cited source did not publish the specific indicators described as Uber-specific network indicators.
Why this matters now
Fireworks’ new separation of two incidents changes assurance requirements. Customers cannot rely on a single containment narrative or one credential-rotation event: the first chain involved internal Fireworks credentials and customer evaluation-job secrets, while the second began with a third-party cloud provider and reached development and production systems. Assurance must identify which chain affected which data and which remediation applies.
Uber’s notice demonstrates a common control failure in delegated cloud access. The environment was under Uber’s cloud-security management and had alerting enabled, yet valid provider credentials and permitted data flows did not trigger those controls. This is a governance and detection-design problem, not merely a supplier patching problem. Enterprises need telemetry that distinguishes expected supplier identity from expected supplier behaviour.
The affected Uber data is durable identity material rather than a revocable password. Driver’s-licence images, licence numbers, names, addresses and dates of birth create fraud and impersonation risk that persists after credential rotation. The updated population also changes legal, customer-support and financial planning for the affected organisation.
The decision for security leaders
Require separate assurance for the two Fireworks incidents. The first needs evidence covering internal credential misuse, exposed evaluation secrets and downstream key use. The second needs evidence covering the cloud-provider breach, development-machine access, production reach and affected stored data. A consolidated statement that the platform is remediated does not prove both chains are closed for an individual customer.
Treat every Fireworks-held secret as a path into another provider until its issuer confirms rotation and log review. Removing a secret from Fireworks is not revocation. Cloud, source-control, model-provider and data-platform owners must validate usage at the systems that accepted those credentials and document whether unauthorised reads, downloads, changes or charges occurred.
Reassess supplier access monitoring. Valid credentials should not exempt activity from behavioural controls. Alerting should incorporate data volume, object sensitivity, new source context, unusual query patterns and deviation from the supplier’s contracted workflow. For durable identity documents, fund individual support and fraud monitoring separately from technical credential containment.
Evidence of closure
- Fireworks’ notice identifies which incident affected the organisation and names the data involved.
- Every exposed credential is revoked at its issuing provider.
- Provider logs record an approved disposition for each exposed credential.
- Regulatory filing reconciliation produces one approved affected-person and data-category record.、Supplier-access detections alert on abnormal data access despite valid credentials.
The Security.io assessment
The new Fireworks update improves causal separation but increases assurance complexity. It confirms that customers may have exposure from different access paths, different dates and different data stores. Enterprises should reject a single generic closure letter unless it states which incident applied, what customer data was accessed, what credentials were involved and what evidence supports containment.
Uber’s disclosure is especially instructive because the cloud environment was managed and monitored, yet the attacker used a trusted supplier identity and allowed data flows. The lesson is not that alerting was absent; it is that identity validity was allowed to substitute for behavioural legitimacy. Supplier-access architectures should be tested against that exact failure mode.
The 149,346-person figure materially changes the downstream impact assessment, but it does not establish the final global scope of Fireworks customers or all Uber jurisdictions. Regulators may publish additional counts or categories. Technical closure, affected-person reconciliation and legal notification should therefore remain separate tracked decisions with independent evidence.
Questions for the morning meeting
- Which Fireworks datasets, models, secrets or billing records belonged to our organisation?
- Could any Fireworks-held credential access systems beyond the platform?
- Do our third-party cloud environments alert on valid supplier credentials crossing expected data boundaries?
- Have legal and privacy teams reconciled all affected jurisdictions and notification counts?