Remote Spectre Attack on Cloudflare Workers Leaked JWT Across Co-Located Isolates

Researchers disclosed a remote Spectre attack against Cloudflare Workers that extracted a JSON Web Token from a co-located Worker running in Cloudflare’s production environment at up to 12 bits per second. The attack targeted separate V8 isolates sharing the same process, using WebSocket communications as a remote timing source and long-lived Durable Objects to bypass assumptions in Cloudflare’s Dynamic Process Isolation (DyPrIs) design. Testing reportedly succeeded on Linux servers with AMD EPYC Zen 2 and Zen 3 processors, with lower system load improving leakage rates while slower exfiltration remained possible under heavier load.

Cloudflare said the research did not access customer data and that it found no signs of active exploitation during the past three years, but acknowledged limitations in its production isolation model and detailed hardening steps. The company said it mitigated the issue by improving DyPrIs, integrating the V8 Sandbox, and deploying MPK-based in-process isolation to better contain speculative-execution attacks in multi-tenant Worker environments. The disclosure highlights that speculative side-channel risks can persist even in sandboxed serverless platforms when attacker and victim workloads are co-located within the same process.