Dynamic Process Isolation for Edge Compute Security

1 post · Jul 2020 – Aug 2026

What this trail covers

Cloudflare Workers has evolved its security model to address side-channel attacks like Spectre and improve the isolation of customer code. This includes leveraging hardware-assisted memory protection keys to isolate V8 heaps between isolates, achieving a 92% protection rate against data leakage. Additionally, the V8 sandbox, a software-based security boundary, is being adopted to prevent memory corruption attacks from escalating. This involves using compressed pointers and removing pointers from the heap to thwart buffer overflow and other memory corruption vulnerabilities, aiming for 100% protection against certain attack vectors.

All Cloudflare trails

Aug 2026A revisit of remote Spectre attacks on Cloudflare WorkersThis post details a revisit of remote Spectre attacks on Cloudflare Workers, building upon previous research and the implementation of Dynamic Process Isolation (DyPrIs). The research uncovered a limitation in the DyPrIs implementation, allowing for a reliable remote Spectre attack that leaked up to 12 bit/s with 99% accuracy in the production environment. This led to improvements in DyPrIs, integration of the V8 Sandbox, and the addition of an in-process isolation mechanism to further mitigate memory disclosure risks. The post also describes the attack primitives, including the speculative type confusion Spectre gadget and signal amplification techniques using PLRU cache replacement policies, and the use of a remote timer for measurements.
Sep 2025Safe in the sandbox: security hardening for Cloudflare WorkersThis post details Cloudflare's implementation of hardware-assisted memory protection keys (PKU) within the V8 runtime for Cloudflare Workers. It explains how random keys are assigned to each V8 isolate's heap data, preventing cross-isolate memory reads and achieving a 92% protection rate. It also introduces the adoption of the V8 sandbox, a software-based security mitigation that removes pointers from the heap to prevent memory corruption attacks from escalating. The post explains memory corruption attacks with concrete examples like buffer overflows and details how V8's compressed pointers (32-bit offsets) contribute to the sandbox's effectiveness by reducing the attack surface.
Oct 2021Dynamic Process Isolation: Research by Cloudflare and TU GrazThis post details research conducted in collaboration with TU Graz on Spectre attacks against Cloudflare Workers. It describes the development of a working Spectre variant 1 attack that could leak memory at a rate of 120 bits per hour under ideal test conditions. More importantly, it introduces and explains the 'Dynamic Process Isolation' defense mechanism, which uses hardware performance counters to detect branch mispredictions indicative of an attack. Upon detection, the Worker is moved to a separate OS process. This defense has been deployed in production and has proven effective against various Spectre variants with a low false positive rate.
Jul 2020Mitigating Spectre and Other Security Threats: The Cloudflare Workers Security ModelThis post details Cloudflare Workers' security architecture, focusing on isolation mechanisms like V8 isolates, process-level sandboxing, and Linux namespaces/seccomp. It specifically addresses V8 bugs and the 'patch gap', and introduces Dynamic Process Isolation as a defense against Spectre by moving suspicious workers to separate processes. The post also outlines API design principles for secure communication and outbound request mediation.
Advertisement