Security1 distinct publisher2 min readUpdated
The kexec fires from LG's own startup.sh after LG's signed kernel is already up. Kernel lockdown, the thing that closes that door, merged at 5.4. The set runs 4.4.3.
The Watch · Security desk
Compiled by The WatchSomething wrong?How this is made
kexec is ordinary plumbing. It is a syscall that loads a second kernel from inside the running one and jumps straight into it, skipping firmware and POST entirely [6]. Enterprise distros carry it because kdump needs it: when a box panics you kexec into a small capture kernel that writes the dead kernel's memory to disk, and that only works because skipping firmware means RAM is never wiped [7]. It is also how a large server reboots without waiting for firmware to count its own DIMMs, and it is the basis for kexec-based bootloaders such as Petitboot and LinuxBoot, where Linux's own drivers and network stack stand in for whatever the vendor shipped [8].
The part worth understanding is why the same move fails on a current Secure Boot machine. It used to work exactly this way. The old `kexec_load` accepts an in-memory image with no integrity checking whatsoever, so root holding CAP_SYS_BOOT could stage arbitrary bytes and enter ring 0, walking around the signed chain rather than attacking it [9]. Kernel lockdown, Matthew Garrett's work, merged upstream in 5.4 to shut that off [10]. Boot a mainstream distro with Secure Boot on today and the kernel enters lockdown, lockdown disables `kexec_load` outright, and the replacement `kexec_file_load` takes file descriptors and runs the kernel's own signature check, so only a validly signed image loads [11]. kdump keeps working because the capture kernel is signed [12].
That threshold sits above the version this television runs, and there is no toggle, command line or build flag that changes the arithmetic: the mitigation is not in that tree to configure [14]. LG's verification carries the chain up to the signed kernel and then stops caring [13], and after that point root and kexec are the same privilege.
So the honest accounting of what the signing chain still does for the owner: it covers everything before the kernel, which is why a bad build costs a reboot rather than a panel. What it never covered was the kernel that is actually executing, and the write-up's own framing is the correct one, according to Paul Asadoorian on PSW #940 - verified boot is a statement about how you started, not about what you are running now [5].
Which leaves the practical blocker on this project as a timer rather than a cryptographic one. Nothing in the signing chain is stopping a custom kernel on this hardware; the SoC watchdog is [15]. Whoever works out how to keep that watchdog fed will have a usable custom-kernel television, and they will not have needed a single key to do it.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
The LG TV custom kernel work is not a secure boot bypass and does not claim to be: LG's signed bootloader, ATF and OP-TEE are never touched. The jump happens from the stock startup.sh after LG's own signed kernel has already booted.
LG's secure boot verifies the chain up to the signed kernel and then stops caring; once that kernel is up and you have root, kexec is right there.
The custom kernel runs entirely in RAM, so a power cycle puts the set back to stock and the method cannot brick the television.
The project is labeled work-in-progress: the custom kernel reaches userspace, but the SoC watchdog resets the TV after 10 to 21 seconds.
Paul's take on PSW #940: verified boot is a statement about how you started, not about what you are running now, and the author deserves credit for saying plainly that they sidestepped secure boot rather than broke it.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Internally coherent single-source explainer, no independent corroboration
The technical account is specific and self-consistent (startup.sh kexec, RAM-only execution, 10-21s watchdog reset, 4.4.3 vs lockdown at 5.4) and the version arithmetic is checkable, but the cluster contains exactly one item, a podcast segment, with no link to the project's code or write-up, no vendor confirmation, and no second publisher.
Pre-production hobby result against a widely deployed kernel facility
The only observed deployment is the work-in-progress project itself, which is RAM-only, reverts on power cycle and survives 10-21 seconds; the source reports no users, releases, or affected-device counts. The one broad adoption fact is that kexec itself ships in every enterprise distro, which is background rather than adoption of this result.
Framing runs cooler than the headline event
The coverage actively de-escalates: it denies a secure boot bypass, credits the author for saying they sidestepped rather than broke it, and stresses that the result is non-persistent and watchdog-limited. Its claims therefore sit slightly below the underlying technical substance, which does establish that a shipped device's chain enforces nothing after the signed kernel.
Security-media attention incentive, running against the sensational read
The item is a security-industry podcast segment whose general incentive is audience attention on hack stories, yet it argues the less clickable position and names no product, sponsor, or vendor it stands to benefit. No sponsorship, vendor relationship, or researcher affiliation is disclosed in the supplied material, so only this weak, self-limiting incentive is observable.
Moderate: mechanism claims plausible and consistent, breadth unverified
Confidence is limited by single-publisher sourcing and absence of the primary project artifact, but raised by the specificity and internal consistency of the technical chain and by the fact that the central conclusion depends only on two version numbers rather than on contested testing.
security
The bottleneck moved: 622 CVEs in July, and no one left to write up the fixes1 distinct publisher
product
Linux 7.2 ships cache-aware scheduling, and Torvalds calls AI bug reports the new normal1 distinct publisher
security
Hugging Face breach ran 69 days: a containment failure, not a rogue-agent flash1 distinct publisher
build
The kernel's Assisted-by trailer makes AI help a signed line, not a rumour1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 20, 2026