This report describes how an AI-assisted researcher exploited writable driver interfaces to escalate from a browser foothold on a Samsung smart TV to root access.
The work shows how a small mistake in a vendor driver can become a full device compromise when the attacker can combine source review, live testing, and careful privilege escalation.
The test began with code execution inside the TV’s browser application, not with a full device exploit.
The team then gave the AI a realistic environment: a controller machine for building ARM binaries, a shell session on the TV, and the matching Samsung firmware source so it could compare code with the live system.
Samsung’s execution restrictions also mattered, so unsigned programs had to be launched from memory rather than from disk.
.webp)
The goal was simple: move from browser-level access to root on the device. To make that possible, the AI had to enumerate the target, identify reachable attack paths, validate a physical-memory primitive, and keep adapting its tooling until the final compromise worked.
Writable Driver Flaw
The key weakness sat in Samsung’s ntksys driver interface, which was exposed with world-writable permissions.
The driver accepted user-controlled physical address and size values, stored them in a table, and later mapped the selected physical page back into user space via mmap. That design effectively gave an unprivileged process a path to raw physical memory access.
.webp)
A second interface, ntkhdma, made it easier to demonstrate the exploit because it leaked the physical address of a DMA buffer.
That leak was not the main bug, but it gave the AI a known page to test read and write access before moving on to more sensitive memory. Once the mapping worked, the researcher had confirmation that the driver could be abused as a physical-memory primitive.
Root Compromise
With physical memory access in hand, the remaining task was to identify the correct kernel data to modify. The browser process was already running as an unprivileged user, so its credential structure in memory became the target.
By scanning the RAM ranges recovered from the system’s boot parameters, the AI found matching credential objects, patched the fields, and launched a root shell.
According to Calif research, the final result was a full privilege escalation from the browser context to uid=0 (root).
In practical terms, that means the TV no longer treated the browser as a restricted app; it treated it as the most powerful process on the device.
This research is notable because it shows that AI can do more than summarize bugs. Given a realistic foothold, source code, and a working lab setup, it can chain discovery, validation, and exploitation into a working root compromise on real hardware.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.
