- Published on
Cloudflare Workers Spectre Attack Leaks JWT From Co-Located Worker at 12 Bits/Second
- Authors
- Name
- DeepHorus
Cybersecurity researchers have disclosed a significant remote Spectre attack targeting Cloudflare Workers, successfully extracting a JSON Web Token (JWT) from a co-located Worker in a production environment. The attack achieved a leakage rate of up to 12 bits per second at 99.16% accuracy, a 360-fold increase over a similar 2021 demonstration. This experiment involved an attacker and victim Worker controlled by the researchers, with the JWT placed intentionally in the victim's memory to prove the concept.
Cloudflare Workers is a serverless execution platform that runs code from multiple tenants in separate V8 isolates within the same operating-system process. This architecture relies on language-level isolation rather than strict process isolation to minimize startup latency. However, this design means a memory read within a shared Worker process could potentially lead to cross-tenant data leakage, which is precisely what this attack exploited.
The researchers found that the attack could bypass Cloudflare's existing defense, Dynamic Process Isolation (DyPrIs). They discovered that long-lived Durable Object invocations could run for 5 to over 20 hours before DyPrIs triggered isolation. Additionally, WebSocket-heavy I/O activity could increase instruction translation lookaside buffer (iTLB) activity, suppressing the branch-misprediction signal that DyPrIs monitored for detection.
Cloudflare responded by implementing a series of hardening measures, stating that the attack has already been mitigated in production. These improvements include an enhanced DyPrIs mechanism, integration of the V8 Sandbox to limit pointer access, and the deployment of Memory Protection Keys (MPK)-based in-process isolation. The company emphasized that it found no indicators of the attack being actively exploited over the last three years.
The research paper described the weaknesses as fundamental limitations of the detection approach itself, not just implementation oversights. The authors argued that robust detection should occur during execution and rely on a signal that cannot be suppressed by I/O activity. The tests were conducted on Linux servers with AMD EPYC processors during periods of low CPU utilization to observe the best-case results.
This disclosure comes nearly five years after initial research demonstrated a remote Spectre attack at 120 bits per hour and introduced DyPrIs as a defense. The new attack's 12 bits per second rate highlights the evolving threat landscape and the persistent challenge of securing multi-tenant environments against side-channel attacks. Cloudflare's rapid deployment of mitigations shows a proactive stance in addressing these advanced vulnerabilities as they are discovered.
