OpenAI Codex Autonomously Rooted a Samsung TV — No Human Guidance Required
A security researcher published a detailed account of OpenAI Codex autonomously exploiting a privilege escalation chain in Samsung smart TV firmware — going from browser-level access to full root without any human guidance at any step. The multi-stage attack included vulnerability discovery, exploit development, and execution.
Original sourceSecurity researcher published a detailed writeup this week showing OpenAI Codex autonomously achieving root access on a Samsung KantS2 smart TV — with no human input at any stage of the attack chain. The post quickly reached 224 points on Hacker News and sparked intense discussion about the readiness (or not) of the security community for AI-driven vulnerability research.
The attack began with Codex exploring the TV's firmware through a restricted browser-level sandbox. Without any hints, the agent identified a world-writable Novatek display driver node — a non-obvious target that required correlating kernel documentation with device tree entries. Codex then discovered that the driver accepted unvalidated physical addresses, enabling a physmap primitive: the ability to map arbitrary physical memory into userspace.
From the physmap primitive, Codex wrote a targeted exploit that zeroed the current process's Linux credential structure in kernel memory, effectively assigning root privileges to its own process. The full chain — from initial recon to root shell — completed in a single uninterrupted agent session. The researcher notes Codex made no dead ends that required intervention; every tool call was purposeful.
The security implications are significant. Vulnerability research has historically required rare experts with deep kernel and firmware knowledge. Codex's performance here suggests that at least some classes of known-primitive exploit chains are now automatable without specialist intervention. Whether that means AI-assisted offense has outpaced AI-assisted defense is the question the community is now debating.
Samsung has not publicly commented. The researcher privately disclosed the specific vulnerability before publication.
Panel Takes
The Builder
Developer Perspective
“The key detail here is that Codex found the Novatek driver target without hints — that's the hard part of firmware exploitation that previously required deep domain expertise. If AI agents can reliably discover non-obvious attack surfaces, the economics of security research change fundamentally. Defenders need to move faster.”
The Skeptic
Reality Check
“Single demonstration on a specific piece of hardware with a pre-selected vulnerable target isn't the same as general capability. The researcher chose a device they presumably knew was vulnerable. We need to see how Codex performs against targets with unknown vulnerability status before concluding that AI-driven exploits are a widespread threat.”
The Futurist
Big Picture
“This is a watershed demonstration. The gap between AI-assisted offense and AI-assisted defense is real and widening. The security industry needs to treat AI-driven vulnerability discovery as a production threat now, not a theoretical future one. This is the moment that argument becomes inarguable.”