The company says affected responses followed requests exceeding 681,280 tokens. Insufficient cache capacity and a flaw in how the SGLang engine handled oversized requests allowed content to cross between customers.
Knostic researcher Gadi Evron describes independent checks ruling out simple hallucinations, with names, invoice information and code among the outputs. Cloudflare says it corrected the configuration and engine before restoring service on 9 September and found no other Workers AI model affected. The figure of 155 counts responses, not an established number of victims.