Copyfail/Dirtyfrag LPE Kernel Patches

The kernel that the dgx spark is running still hasn’t received critical security updates that allow local privilege escalation.

Considering the nature of the agentic workloads on the platform you’d expect this would be a priority.

Can we at least get an official kernel patch more recent than the current mid March build with the proper upstream security fixes?

I’m aware of the module work arounds but I’d like some reassurance the GB10 platform won’t be stuck here on an outdated kernel and an outdated driver train @580

You don’t really need a kernel patch.

$ cat /etc/modprobe.d/lpe-mitigations.conf
install algif_aead /bin/false
install esp4 /bin/false
install esp6 /bin/false
install rxrpc /bin/false

Yes, I know as I said in the post.

Copy Fail requires local access to exploit. Unless you’re an enterprise it likely will never impact you. If an attacker has local access to your machine, you have bigger issues to worry about than a patch for Copy Fail.

There are a number of ways to achieve RCE. For example, an agent running on the Spark might offer the opportunity to achieve RCE by means of prompt injection. From there, chaining the LPE such as Copyfail or Dirtyfrag would be quite fast.

I don’t want to argue for the sake of it but that can be done with a ton of other vulnerabilities and isn’t unique to CopyFail. There could be a lot of “what ifs” that could achieve RCE that apply to many other vulnerabilities. I hope they release a patch to help cover this for folks and I’m not trying to downplay the severity of the vulnerability. I’m just trying to help folks understand that there are a lot of critical vulnerabilities that can achieve root and you can’t expect the Nvidia team to bend over backwards for every one of them. Just have to look at it from a realistic lens based on how long it takes us to get updates for everything else, it might take awhile.

It’s as simple as providing routine kernel patches since they’ve customized the Ubuntu kernel.

At this point we’re two months back and 3 revs back on compute/graphics drivers.

These are a couple of the most severe exploits to date on a platform that runs AI agents and 3rd party code.

If you don’t think it’s an issue, just run every workload on the spark as root.

You’re entitled to your opinion. As I said, not here to argue. Trying to provide an explanation. You have to understand the difference between likelihood of exploitation and severity, because one does not equal the other. EPSS (exploit prediction scoring system) is extremely low. Not going into details on what exactly that metric means, but I do this for part of my daily job. It’s a severe and yet also over hyped vulnerability. If you’re putting this much energy into it, just apply the workaround and wait for the official patch. You said it yourself, they’re 2+ months behind. Have a good day!

Yes, such an overhyped vulnerability that cloudflare had to touch almost every server in their footprint.

How Cloudflare responded to the “Copy Fail” Linux vulnerability

Thank you for your opinion, I’ll let the netsec community know.

Cloudflare deals with public facing infrastructure. Not sure why you’re being so dense. Most folks here are running locally. Muting this thread, you’re being really toxic.

Agentic workflows and 3rd party code are expected to run in a sandboxed/isolated environment.

These exploits directly nullify that isolation.

Whatever workflow you’re using on the spark is apparently hobby level enough that it can run as root without a second thought, but some people have to actively deal with compliance.

Here’s an image for the visual learners:

Wait, you’re saying that letting strangers or bots execute arbitrary commands on my DGX Spark is a bad idea?

What happened to trusting people in this world? shaking my head

Overall, this is a overall Linux issue, not limited to the DGX spark. Ubuntu will release a kernel update when it’s available for their LTS versions, before that I don’t think we can expect any update in relation to this. I’m not even sure that the kernel has a proper patch for this yet?

The ubuntu patch has been released since the beginning of the week.

Oh, then we should get it “soon”.

Either way, i believe the ones who would have issues with this bug on their DGX spark is already have a poor sense of judgement for what to run on it, that’s just my 2 cents.

Glad someone else said it.

I think this is a little narrow viewed considering the overall pattern of delayed software support/releases for the gb10/spark platform.

LPEs regardless of how the system is implemented and what workflow is running are a risk, we can go back and forth about decisions that are in or out of anyone’s control but ultimately exploits need to be patched. If not for functional/security requirements then at least for compliance/audit reasons. When nvidia effectively takes over the kernel development for ‘dgx os’ away from upstream Ubuntu kernels the responsibility is now on them to deliver timely security patches.

LPEs today, RCE tomorrow, the trend is ‘spark gets significantly delayed attention’

The trend of this entire community running vibe coded personal projects and openclaw/nemoclaw means that this is an incredibly vulnerable platform for any kind of late breaking/embargo broken exploit .

Now that the serious debate is out of the way, here’s a light hearted meme that is not intended to mock anyone.

I can get behind spark gets significantly delayed attention as a cause for concern. Should/will they release a patch? Yeah. Will it be soon? Probably not.

I don’t know how many are or aren’t running the agents themselves on their DGX spark.

But as you mentioned in an earlier post, you should have the isolation.
To me that means that the LLM (vllm for me at least) and my agent (pi, opencode, claude code, hermes, openclaw, custom agents, etc) means to run them in it’s own separate machine (probably virtual, and i do not count docker in here) over a no trust network.

If you’re not having clear separation between all parts, then you’re always going to be having these issues, as soon as one thing is compromised/faulty the entire network is at risk.

So is it bad that we don’t have things patched day on? Yes.
Is the vast majority of companies or people running updates day 1 anyway? No, and either can be a risk.

The likelihood that Nvidia would ever take over the kernel development is very low (doesn’t seem to be even remotely close to needs or financially reasonable, custom patches maybe… similar to many distros)

So my recommendation for anyone running things things on the Spark directly, invest in a Raspberry pi or similar to run the agent, and have the actual LLM running on the spark, separate.

@trystan1 Side note: I love those images! keep them coming <3

I can guarantee that emerging, embargo broken, deterministic exploits are patched immediately by large companies due to the risk.

I’m assuming you’re alluding to supply chain attacks ala npm, that’s why the vendors themselves typically now have internal betas/opt ins before production releases but this happens within days, not weeks or longer as is seen in this example.

The fact remains, outdated kernels on any machine are a risk. Copyfail is patched upstream and has been for weeks, the distributions were caught off guard specifically because of how the communication was handled so they ‘only’ took a few days. But every major distribution including ubuntu has kernel updates.

Nothing from nvidia, it’s a problem and the track record/reputation for delays causes concern and reassurances that at least SOMETHING is in the pipeline was what I was attempting to confirm.