# Spectre Attack on Cloudflare Workers Leaks JWT at 12 Bits/Second Co-located Worker in same process leaked sensitive data at unprecedented rate. - **Source:** [The Hacker News](https://brifing.si/source/f944ec2b-8180-47bc-94de-57085b5a6563) - **Published:** 2026-08-19 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/63e1396e-9cfc-4ece-9552-a9179c42a2ed ## Scale of the Spectre Attack Researchers demonstrated a remote Spectre attack against Cloudflare Workers, successfully exfiltrating JSON Web Tokens (JWT) from a co-located Worker within the same process. The attack achieved data leakage at 12 bits per second - 360 times faster than a similar attack demonstrated in 2021. The experiment used attacker and victim Workers controlled by the researchers. ## Cloudflare's Mitigation Measures Cloudflare stated it has mitigated the attack in production by improving Dynamic Process Isolation (DyPrIs), integrating V8 Sandbox, and deploying Memory Protection Keys (MPK)-based isolation. The company found no evidence of active exploitation in the past three years. However, researchers proved DyPrIs implementation in production was insufficient against this attack vector. ## Architectural Vulnerability Cloudflare Workers runs multi-tenant code in separate V8 isolates within the same OS process. While this language-level isolation reduces startup latency, memory reads within a shared Worker process can lead to cross-tenant leakage. The attack requires both attacker and victim Workers to be co-located in separate V8 isolates within the same Worker process. ## Timing Sources Exploitation Workers restrict local timing sources during CPU execution. However, researchers discovered WebSocket communications could provide remote timing, and Durable Objects could keep a Worker isolate alive for 5-20 hours. While DyPrIs isolates suspicious scripts post-invocation, long-lived Durable Object invocations continue running before isolation occurs. ## Detection Mechanism Flaws Heavy WebSocket I/O activity increases instruction translation lookaside buffer (iTLB) activity, reducing DyPrIs' normalized branch-misprediction signal below detection threshold. Cloudflare considers this a DyPrIs implementation limitation, while researchers argue it reflects fundamental constraints of the detection approach. Robust detection should occur during execution using I/O-resistant signals. ## Testing Environment and Results Production tests were conducted on Linux servers with AMD EPYC Zen 2/Zen 3 processors during low-utilization periods (10-25% CPU). While higher system load reduced leakage rate, slower attacks remained feasible. The attack achieved 12 bits/second leakage at 99.16% accuracy, compared to just 2 bits/minute in 2021's attack. ## Historical Context Five years prior, Cloudflare and TU Graz published research demonstrating a remote Spectre attack against Workers at 120 bits/hour, introducing DyPrIs as defense. The 2021 paper reported 0.61% false-positive rate, concluding DyPrIs provided security guarantees equivalent to strict process isolation against then-known Spectre variants. ## Industry Implications These findings necessitate reevaluation of security assumptions in shared infrastructure models. Cloudflare's September 2025 hardening measures include random MPK assignment (trapping ~92% cross-isolate accesses) and stricter rotating layout to address same-key scenarios. The research underscores persistent challenges in securing multi-tenant serverless environments against sophisticated timing attacks. ## Sources covering this story - https://thehackernews.com/2026/08/cloudflare-workers-spectre-attack-leaks.html — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._