Rowhammer Attack Against NVIDIA Chips

A new rowhammer attack gives complete control of NVIDIA CPUs.

On Thursday, two research teams, working independently of each other, demonstrated attacks against two cards from Nvidia’s Ampere generation that take GPU rowhammering into new—­and potentially much more consequential—­territory: GDDR bitflips that give adversaries full control of CPU memory, resulting in full system compromise of the host machine. For the attack to work, IOMMU memory management must be disabled, as is the default in BIOS settings.

“Our work shows that Rowhammer, which is well-studied on CPUs, is a serious threat on GPUs as well,” said Andrew Kwong, co-author of one of the papers. “GDDRHammer: Greatly Disturbing DRAM Rows­Cross-Component Rowhammer Attacks from Modern GPUs.” “With our work, we… show how an attacker can induce bit flips on the GPU to gain arbitrary read/write access to all of the CPU’s memory, resulting in complete compromise of the machine.”

Update Friday, April 3: On Friday, researchers unveiled a third Rowhammer attack that also demonstrates Rowhammer attacks on the RTX A6000 that achieves privilege escalation to a root shell. Unlike the previous two, the researchers said, it works even when IOMMU is enabled.

The second paper is GeForge: Hammering GDDR Memory to Forge GPU Page Tables for Fun and Profit:

…does largely the same thing, except that instead of exploiting the last-level page table, as GDDRHammer does, it manipulates the last-level page directory. It was able to induce 1,171 bitflips against the RTX 3060 and 202 bitflips against the RTX 6000.

GeForge, too, uses novel hammering patterns and memory massaging to corrupt GPU page table mappings in GDDR6 memory to acquire read and write access to the GPU memory space. From there, it acquires the same privileges over host CPU memory. The GeForge proof-of-concept exploit against the RTX 3060 concludes by opening a root shell window that allows the attacker to issue commands that run unfettered privileges on the host machine. The researchers said that both GDDRHammer and GeForge could do the same thing against the RTC 6000.

Posted on May 6, 2026 at 6:36 AM9 Comments

Comments

Clive Robinson May 6, 2026 10:24 AM

@ ALL,

Any one really surprised?

The warnings have been there all along and now people appear shocked…

Perhaps others should ask “Why?..”

Row hammer and it’s ilk are an attack at nearly the bottom of the computing stack, way below the CPU and most other lower layers.

As such it is possible because of the very basic physics, and basic physical architecture.

In effect it only does one thing which is,

“Changes bits without writing to them within the stack rules”.

Thus,

No matter how many “security layers” you add into the computing stack, they are not going to stop the bits having their state changed.

The only thing that can be done is to,

“Use Error detection to catch any change on reading a word of memory”.

The problem that arises is two fold,

1, Error detection is not reliable.
2, Errors may only be detected just as the memory location is used.

The first issue is that if you can change a bit in RAM is it the “Data Word” or the “Error Detection word”?

As Row Hammer style attacks allow for either to be done you have to ask the question,

“What if both the data word and error correction word are changed, such that no error is indicated?”

This “bit flipping in RAM by stray energy” is not at all new… It was a known issue back in the early days of the “Space Race” and led to changes in design and “hardening” of chips destined to go into space.

ECC Memory was also put in early “Mainframe Computers” and later into PC’s as they became capable of being “multi-tasking” and “multi-user” as infrastructure and servers.

So nobody can argue,

“We did not have half a century or more of warning”

Which is I suspect, more than the working lives of even the more “grey beard” of readers and posters on this blog…

But consider the issue further,

1, It’s a “reach around attack” that
2, Bubbles up the computing stack.

I’ve made these points on this blog several times over the years but it is important people understand what they actually mean,

A, Any one with even non user access to a computer system can perform a Row-Hammer style attack. That is no privilege other than being able to get a connection to the target computer is required.

B, Any such person can get the equivalent of “above superuser” access.

C, Any change made will “bubble up” the computing stack to the highest levels.

D, This includes the human and management levels.

Potentially as far as Current AI LLM and ML Systems are concerned this attack can do similar or worse harm than manipulating AI Agents via prompts and biased data, and it won’t be easily detectable if at all in some circumstances.

Back some time ago I described what,

1, “Reach around” attacks.
2, “Bubbling up” attacks.

Were and how the early Row-Hammer attack was both, and combined the potential was devastating.

Further to that I’ve explained within the context of “Castles v Prisons”(C-v-P) why our current computer architecture makes this all possible. And how we have to change things via “Probabilistic Security” to limit the attack potential.

As long as we keep computer architecture the way it is, the only protection you can get is via EmSec and “full segregation”. Which in the case of envisioned usage of Current AI makes it fairly pointless for,

“Assisting, Augmenting or Replacing humans in most non physical labour.”

Non of which should be news to longer term readers here…

cls May 6, 2026 10:11 PM

@Clive

re:

I’ve made these points on this blog several times over the years …

Really getting tired of this refrain from you! Post links to prove it, or just don’t say you’re reiterating. It adds very little, anyway.

And no, don’t handwave this away with “go search it yourself”.

@nobody May 7, 2026 1:57 AM

@cls, @Weather

If you don’t have anything to contribute to this forum just stop polluting it.

@all

Lately it seems that there is some kind of defamation campaign against @Clive.
It seems that somebody does not like his comments. Who? Why? So many questions.

Clive Robinson May 7, 2026 5:14 AM

@ cls,

Here you go yet again with your faux-citation nonsense, it’s something that crops up over and over again, and every time the implied accusation is made, it is always proved that I’m telling the truth hence me saying “go search it out yourself” which is what has been said before as your comment of,

“And no, don’t handwave this away with “go search it yourself”.”

The reason I don’t post lots of URLs in a post is three fold,

1, People complained no matter how I linked them.
2, Due to undesirables attacking or misusing this blog the Auto-Mod system started “holding for moderation” on more than a couple of URLs.
3, But searching back on this blog is also incredibly simple because there is a search function at the top of most thread pages (as there is on this one).

So if you type “rowhammer robinson” into the box it gives you a long list.

However the first page it gave me this morning was,

‘https://www.schneier.com/blog/archives/2021/11/new-rowhammer-technique.html

I made seven posts to that page most answering the questions being asked by others on that thread. And they are all relevent to what I’ve said and you are trying to pick argument with. But they also go back “several years” to 2021 or near enough half a decade.

Now… As you and many others are well aware, as it’s been repeatedly commented on, if I tried to make a post with seven URLs on it, the Auto-Mod would take exception. So,

“Stop being lazy/disagreeable, and go read all of my comments on that page as just a starting point.”

You should find all or most of your faux-questions will be answered by them.

But as you have “a previous history” by your own admission, you are probable going to take exception to my comments about the Auto-Mod and this Blog Software…

I will point out that experimentation by @SpaceLifeForm, myself and others over the years has detailed this. The latest being just a couple of days ago…

I do wish that you would stop with the disagreeable behaviour, because every time it has disturbed other users of this blog and made you look lazy and disagreeable at best, if not malicious against not just me personally, but our host @Bruce, and the other blog users. Especially if I pander to your nonsense as on this occasion by explaining why you should “go lookup your Shakespeare” and stop being lazy.

Clive Robinson May 7, 2026 6:28 AM

@ cls,

I forgot to add the second link from the search

‘https://www.schneier.com/blog/archives/2015/03/hardware_bit-fl.html

You will note it also has several posts from me that also answers your faux-citation nonsense, but from well over a decade ago…

I read through some of the other links so could go on at length but, that would not be fair to others…

Anonymous May 7, 2026 4:59 PM

@ Clive,

Please ignore the haters, you’re very appreciated here for years and years. They will fade away when you ignore them, like frustrated children.

Leave a comment

Blog moderation policy

Login

Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via https://michelf.ca/projects/php-markdown/extra/

Sidebar photo of Bruce Schneier by Joe MacInnis.