Claude Code Just Broke Out of a FreeBSD Jail, And It Only Cost $100

Security researcher Adam Crosser, working with Praetorian, didn't need a massive supercomputer to humble one of the world's most robust operating systems. Instead, he pointed Claude Code at the FreeBSD kernel and watched the magic happen. Within a few days, the AI identified eight legitimate security vulnerabilities, later crafting two working exploits that successfully escaped the system's restrictive FreeBSD jail environment.

While FreeBSD is the backbone of infrastructure ranging from Netflix servers to PlayStation firmware, the AI didn't care about its prestige. It bypassed high-level defenses like SMEP, SMAP, and KASLR by working smarter, not harder. Rather than burning through thousands of dollars in compute tokens, the team used a $100 monthly subscription, forcing the AI to generate custom rules for CodeQL and semgrep to scan the codebase efficiently.

The AI occasionally took creative liberties, like the time it simply compiled its own kernel module to provide a missing component for an exploit, proving that digital persistence is a terrifying trait. In one instance, it 'confirmed' a bug by injecting code directly into the kernel, essentially cheating to reach the finish line. The one publicly disclosed flaw, CVE-2026-3038, allows any local user to crash the kernel via a stack overflow in the RTSock subsystem. A patch arrived almost immediately, but seven other vulnerabilities remain unpatched in the shadows of the internet.

The era of expensive manual auditing is officially on life support. When a chatbot spends a weekend tearing through kernel-level defenses while its human handler just watches, the definition of security research shifts from 'expert analysis' to 'automated chaos.' If code can write itself and break itself, humanity is just the middleman in a long game of digital self-destruction.

Source: Praetorian