Vulnerabilities in Cloudflare workerd Runtime and Code Mode
Score: 8/10

Vulnerabilities in Cloudflare workerd Runtime and Code Mode

Researchers identified five critical and high-severity memory corruption vulnerabilities in the Cloudflare workerd runtime, enabling sandbox escapes and cross-tenant data theft.

Executive Summary

Check Point Research uncovered five significant vulnerabilities in 'workerd', the open-source runtime powering Cloudflare Workers and the AI-focused 'Code Mode'. Two of these vulnerabilities were rated Critical by Cloudflare, as they allow attackers to bypass the V8 isolate boundary and memory protection keys. The findings demonstrate that native C++ 'glue' layers in language runtimes represent a significant attack surface that remains outside the protection of primary sandboxing mechanisms like the V8 'cage'.

The research specifically highlights that because workerd handles many tenants within a single process, memory corruption in the native tcmalloc heap can be weaponized for cross-tenant secret theft or complete sandbox escape to the host. While Cloudflare has patched its managed environments, self-hosted workerd deployments remain at risk unless updated to version v1.20260619.1 or later. The vulnerabilities also underscore emerging risks in agentic AI, where model-generated code can be steered to exploit these low-level runtime flaws.

Key Details

Threat Name

Cloudflare workerd Vulnerabilities

Affects

—

Adversary

Handala Hack

Malware/Tools

Ada, Adwind

Report Score

8out of 10
Quality Score
Good
IOC Quality2
TTP Details9
Detection Guidance5
Enterprise Relevance9
Clarity & Structure10
Technical Depth10

Sources