Cloudflare Shows a Full Remote Spectre Attack Against Workers

Researchers at Cloudflare, the University of Edinburgh, and TU Graz ran a full remote Spectre attack in the production Workers environment and leaked data from a co-located victim Worker at up to 12 bit/s, with about 99% accuracy. That puts the leak squarely between co-tenants sharing one OS process, the exact arrangement every serverless platform runs on.

Two bypasses made it work. The team got around the platform's timer restrictions with a remote WebSocket timing source, then waited out the DyPrIs detection system by holding Durable Object invocations open for hours. Workers co-tenants untrusted code in a single process, and this attack ran about 360x faster than the 2021 baseline. Side-channel attacks on serverless runtimes look a lot more practical than the industry has been assuming.

You can stand down for now. Cloudflare has already mitigated the attack in production, and its telemetry found no active exploitation over the preceding three years. Nothing for you to patch this week. But if you build on serverless, this is the reminder that your threat model includes the microarchitecture your neighbor's code shares, not just the advertised sandbox boundary. For anything holding secrets, co-tenant isolation just got harder to wave off.