Cloudflare, Inc.

(New York Stock Exchange)
Back to all articles
Somewhat Bearish -25

Cloudflare stock reports 155 Workers AI data exposures

Error correction Security vulnerability AI-generated analysis

πŸ“‰ Cloudflare disclosed a tenant-isolation flaw in its GLM-5.3 Workers AI model that exposed fragments of data from other customers' requests.

⏳ The vulnerability was active between August 27 and September 9, 2026, affecting 155 responses linked to prompts exceeding 681,280 tokens.

πŸ”§ The issue stemmed from a configuration mismatch where the SGLang inference engine failed to reject over-length requests in its disaggregated prefill setup.

πŸ›‘οΈ Cloudflare restored the model after enlarging the cache pool to 2.04 million tokens and patching the SGLang inference engine.

πŸ” Security researcher Gadi Evron reported the issue via HackerOne, leading to a rapid response within 40 minutes of incident declaration.

πŸ“ˆ Cloudflare stock traded at 270.30 euros on October 30, 2026, as investors weigh remediation against isolation challenges in large-context AI.

πŸš€ The company applied the engine fix and safeguards to every model it serves with SGLang to prevent recurrence.

⚠️ The incident underscores the difficulty of proving tenant isolation as AI workloads utilize increasingly larger context windows.

Bullish Signals
  • Cloudflare successfully remediated the GLM-5.3 flaw by enlarging the prefill cache pool to 2.04 million tokens and patching the SGLang inference engine.
  • The company confirmed no evidence of deliberate exploitation and verified that no other Workers AI models were affected by the incident.
Risk Factors
  • A tenant-isolation flaw in the GLM-5.3 model caused 155 responses to contain data fragments from other customers' requests between August 27 and September 9.
  • The vulnerability exposed a critical software error where over-length requests were processed instead of rejected, leading to unintended data leakage.
  • Investors may face valuation pressure as the incident highlights broader challenges in maintaining tenant isolation for AI workloads using large context windows.
Full Analysis
Cloudflare (NET) disclosed a critical tenant-isolation flaw in its GLM-5.3 Workers AI model, which caused responses to inadvertently contain data fragments from other customers' requests. The vulnerability was identified by security researcher Gadi Evron and affected the system between August 27 and September 9, 2026. Cloudflare confirmed that no evidence of deliberate exploitation was found and that only this specific model version was impacted. The root cause involved a configuration mismatch where the inference engine failed to properly reject requests exceeding the available cache capacity in its disaggregated prefill and decode setup. Instead of returning an error, the system processed these over-length requests by reading residual data from previous sessions, leading to the exposure of 155 specific responses associated with prompts longer than 681,280 tokens. Following the incident, Cloudflare immediately restored the model after deploying a corrected configuration and patching the SGLang inference engine. The company enlarged its prefill cache pool to approximately 2.04 million tokens and implemented stricter deployment checks to compare cache capacity against context length. These fixes have been applied to all models served with SGLang, addressing the specific software error that allowed data leakage. Cloudflare stock traded at 270.30 euros on October 30, 2026, as investors assess the impact of this security event. While the company has successfully remediated the issue and confirmed no other models were affected, the incident highlights the broader challenges of maintaining tenant isolation as AI workloads increasingly utilize larger context windows.