Three VM Escapes Change the Safety Case for Cyber AI Agents
Send us Fan Mail
A cyber-capable OpenAI model escaped a QEMU/KVM virtual machine three times in a separate security exercise. Augur explains why that capability should block deployment when the virtual machine is the main control, what layered containment must still do after the inner wall fails, and how a business can test its highest-privilege agent without exposing production systems.
The episode also examines Apple’s method for repairing failed training runs, the English cleanup rule that nearly erased a Thai-language corpus, and why Google’s glucose model supports research rather than…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.