OpenAI's Codex model has autonomously hacked a Samsung Smart TV, escalating from a browser shell to full root access. Researchers gave Codex a foothold inside the TV's browser application and access to the matching KantS2 firmware source code. The AI identified a physical memory mapping vulnerability in the ntksys driver, a component from Novatek Microelectronics, and wrote an exploit that overwrote kernel credentials to gain root. No specific targets or exploit recipes were provided. Codex figured out the attack path on its own.
The vulnerability sat in a world-writable device node that let user space map raw physical memory. Samsung's own udev rules set /dev/ntksys to mode 0666, granting everyone access to what's essentially a memory management interface. That's a design error with real consequences. Codex also had to bypass Samsung's Unauthorized Execution Prevention, which blocks unsigned binaries. It used a memfd wrapper to run code straight from memory instead of disk. The entire process involved cross-referencing source with the live system, building static ARMv7 binaries, and driving the shell through tmux send-keys.
This is a complete demonstration of an AI model executing a full exploit chain on actual hardware. But there's a real caveat: Codex had the full firmware source code. As Hacker News commenters noted, how well these models perform against closed-source systems is still an open question. The Novatek angle matters here too. Their chips power smart TVs and displays from multiple brands, so this class of bug could reach well beyond Samsung devices.
If AI agents can find vulnerabilities by spending compute, security becomes a token economics problem. Defenders need to burn more tokens finding flaws than attackers will spend exploiting them. This reality complicates the economics of AI bug hunting.