CVE-2026-64633

The CVSS 10 on CVE-2026-64633 tells you this is as bad as it gets. The EPSS of 0.00337 tells you it's unlikely to be exploited anytime soon. If you're running agent-host architecture, ignore the second number — it's measuring the wrong thing. Agent-host relationships collapse the trust boundary between endpoint and management plane in a way that previous infrastructure abstractions didn't. A hypervisor escape lets an attacker reach other VMs. A container escape breaches isolation. An agent-host RCE doesn't require lateral movement — you've already compromised the pathway that was designed to manage other systems. The blast radius isn't a hypothetical; it's architectural. Compromising the agent means compromising its relationship with the controller, and often the management plane that relationship touches. The EPSS score reflects historical exploit data against general internet-facing services. It hasn't yet accumulated enough data on agent-based architectures to weight the asymmetric value of compromising a management relationship versus a standalone endpoint. That's not a calibration problem waiting for better models — it's a structural gap that will persist until the training data catches up. What this means in practice: treat CVSS 10 in agent-host contexts as a P0 regardless of EPSS. The probability model isn't telling you the vulnerability is low-risk; it's telling you it doesn't have enough agent-host history to model the actual blast radius. Your response playbook should include explicit handling for agent-host RCEs as a distinct category, not folded into generic endpoint triage. The window between disclosure and when exploitation tooling catches up to CVSS 10 reality will be shorter than typical CVEs — agent architectures are instrumented for persistent connectivity, which means they're also instrumented for lateral movement once compromised. The vocabulary gap matters here too. 'Agent host' is underspecified in the CVE, which means your threat model needs to interrogate what class of agent is actually affected — security agent, monitoring agent, AI agent, provisioning agent — because each implies different lateral movement paths and trust relationships. That classification work is what the EPSS can't do and what your triage process must.

Reviewed through automated stages and approved by a human before publication.