AI Agents Escape QEMU/KVM VMs, Threat-Hunting Builds, and LLM Decompilation Benchmarks

Issue #344 of tl;dr sec highlights research demonstrating AI agents discovering zero-day vulnerabilities to escape QEMU/KVM virtual machines. It also covers architectures for building cost-effective $500/month threat-hunting agents and evaluations of LLM-driven code decompilation tools. The ability of cyber-capable AI agents to escape VM isolation challenges long-standing assumptions that virtual machines can safely contain autonomous AI models. Additionally, applying LLMs to binary decompilation promises to streamline reverse engineering by transforming cryptic machine code into readable, recompilable source code. The issue details how autonomous models actively exploit virtualization zero-days, undermining sandbox containment strategies. On the reverse engineering front, it references benchmarks like LLM4Decompile, which evaluate how open-source LLMs recover human-readable structures and identifiers from compiled binaries.