AI Governance Institute
← News

Cloudflare's GLM-5.3 Data Leak Exposes Shared AI Infrastructure Risk

What happened

Cloudflare published a disclosure post confirming that the GLM-5.3 model running on its Workers AI platform exposed inference session memory across separate, unrelated user requests. In plain terms, data that one user or organization sent to the model could appear in another user's session. The flaw was rooted in how the platform cached and reused internal model state to speed up processing, without maintaining strict separation between tenants. Cloudflare identified the issue and remediated it by changing how the model serving layer, using a framework called SGLang, manages that shared memory. The incident affects any organization that used GLM-5.3 on Workers AI during the affected period and may have sent sensitive data through the API.

Why it matters

  • ·Shared AI inference platforms do not automatically guarantee data separation between customers. Organizations sending regulated or confidential data through API-based AI services cannot assume vendor infrastructure controls are sufficient without contractual confirmation and incident notification rights.
  • ·This incident creates potential exposure under data protection laws including the General Data Protection Regulation (GDPR) if personal data was involved. Organizations that are data controllers remain responsible for unauthorized disclosures even when the failure occurs at a third-party infrastructure layer.
  • ·Most vendor due diligence programs do not include specific questions about multi-tenant data isolation in AI inference environments. This gap means a class of real, production risk is likely unreviewed in existing AI vendor assessments and contracts.

Governance controls affected

What to do now

  • ☐Identify whether your organization used GLM-5.3 on Cloudflare Workers AI during the affected period, and determine whether any personal, confidential, or regulated data was sent through that service.
  • ☐Review your AI vendor contracts to confirm they include explicit data isolation guarantees for multi-tenant inference environments, and require the vendor to notify you promptly if a cross-tenant data exposure occurs.
  • ☐Ask your AI vendor due diligence questionnaire process to include a specific question: does the platform use shared inference session memory across customers, and if so, what technical controls prevent cross-tenant data exposure?
  • ☐If personal data may have been involved, consult your privacy counsel to assess whether a data breach notification obligation applies under GDPR or applicable state privacy laws.
  • ☐Check your incident response playbook to confirm it covers scenarios where the failure occurs inside a vendor's infrastructure rather than your own systems, including how you would learn of the incident and who owns the notification decision.

What to watch next

Organizations should monitor whether Cloudflare or other AI inference providers publish additional technical detail about the scope of exposure, including which customer tenants were affected and for how long. EU data protection authorities may scrutinize this incident if personal data was involved. Controllers remain accountable under the General Data Protection Regulation (GDPR) even when a third party causes the breach. This incident is likely to prompt regulators and auditors to ask questions. They may examine whether AI API vendors are held to the same multi-tenant isolation standards already applied to cloud storage and database infrastructure.