I made a tiny ai bug hunting harness (<4MB) that has everything (except the model obviously). It was designed for pentesting purposes where the tiny size matters to make it more portable between environments.
The intended purpose is not to be used as a worm but it does not take a genius to figure out that with small modifications such a thing could work relatively well - especially if it uses AI keys from compromised targets. Making the agent self-modifiable is relatively straightforward task and in fact I already did that in another project.
Every Windows computer has a small rwkv model on it. Wouldn't be hard at all to get decent cpu performance from a tiny malicious harness, especially one that used the self-evolving skills features and open source models.
Malware is going to be crazy, people aren't ready for the revelation of how insecure and broken things are. Everything is held together by bubblegum, duct tape, and panicked engineers putting out fires.
In the paper they say that the worm uses either existing vulnerabilities that it has been trained on or new published vulnerabilities that it scrapes. 44% claimed success.
The paper is a bit silent on why a such a worm would need an LLM. It seems that brute forcing all known vulnerabilities, script kiddie style on each new machine is about the same.
But apparently that info is too dangerous to release ...
It's not fully described how things work exactly, but apparently it does not transfer entire LLMs as part of the worm. Now that would be interesting :)
> The worm parasitically uses compromised machines to run open-weight large language models (LLMs) to sustain its reasoning, or extend its reach for further attacks.
Thanks for pointing that out. I scanned the paper and found that in their main experiments, they use a shared GPU resource and do not copy LLMs to target machines. Apparently they did other experiments in the ablation study where they did copy LLMs.
So it's even worse than I expected. The intended worm can spread through my thermostat, and when it reaches a GPU host, it can spread even harder. Fun times ahead.
I wonder if gamma ray memory corruption will induce a sort of mutation and selection effect on non-ecc-memory hosts which will make the worms effectively evolve.
I think an approach could be to use some engineered security issue or however people build botnets, and give it some AI llm that is small and minimal but comes with instructions to download models from hugging face, and some other minimal prompts and descriptions of tools. Then it could use this to grow in infected computers and try find more capable and vulnerable computers to run better capable models and also devise some minimal communication between the different points of the botnet. Perhaps set itself a goal to dominate the biggest amount of compute and have some other goal. Would be curious to see what happens.
If you infect a machine with GPU enough to run the localLLM needed to steal another machine, you can let it burn tokens all day for free because whoever you stole the first one from will pay the electric bill.
Doesn't Neal Stephenson's Snow Crash have a similar idea? IIRC, a computer infects human brains via language and sound.
In the 2004 Battlestar Galactica series, the explanation for why the Galactica was the only ship that survived a massive Cylon attack seems more and more likely. The ship was old and wasn't fully connected to the human's command and control systems and so the Cylon virus couldn't reach it.
ANY online device? Even assuming AI can find vulnerabilities in every operating system, there's no indication that this is actually true beyond a "here's how it could work"
This is the same nonsense that lead to article saying researchers had created a wormhole when all they had done was draw one.
I have a microcontroller with an ROM disk (i.e., physically read only). You're telling me that an AI can find a way around the physics of not being able to mutate ROM and exploit it?
I mean, if it's online it has a network/wireless card and a TCP stack along with at least some amount of RAM, so yea, in theory unless the programming is perfect it could be exploited. Now, it's not going to be used to run AI, but could very well get used in a DDOS or something like that.
Did people doubt that this was theoretically possible? Seems self-evident to me. The interesting thing will be seeing it in the real world rather than in a controlled environment where they deliberately made all devices on the network have a known vulnerability.
There’s a difference between speculation and measurement, especially since you’d have people making arguments like saying that open models aren’t powerful/fast enough to work. Demonstrating this is a useful warning to everyone (most of the industry) who’s been slacking on internal defenses because they don’t think a well-resourced attacker will target them.
Honestly with some of the denialists here a terminator could kick down their door with lazgun in hand and they'd still tell you that AI can't do that.
And for the people that think that alignment is stupid, not training your AI to think twice about writing self spreading worms is a recipe for disaster after someone gets a token stealing, resource grabbing worm going.
Yeah, lab leak is hard enough to contain with human viruses, but labs have well established protocols to prevent it happening.
Computing doesn't have good protocols except for air-gapping, we really just have lots of layers of best-effort detection, and billions of devices which mix data and instruction often in a careless fashion.
I used to not believe in the dangers of AI or the risk of internet-collapse from "rogue AI", but a genuine self-mutating virus could genuinely take down the internet and need an entirely new separate net. ( Or we'd discover if the current backbone actually has the power to break encryption to stop it. )
And this time, you can bet any new internet would be corporation captured. CompuServe and AOL failed because of the open internet, but we're a very different world now, governments would support the corporation led locked-down approaches for "safety".
I don't for a second believe the capability is actually there yet, but it's no longer unthinkable that such a thing could be created in a lab within a decade. Once out in the wild, there's a lot of idle compute out there to harness for self-improvement and spreading.
Obvious pattern of using ai to replace human reasoning in a proven methodology of malware distribution, C&C, and network infiltration obviously possible, say researchers.
Researchers use AI to create the torment nexus using commodity hardware, demonstrating the very real threat that AI could enable attackers to create torment nexus nodes using commodity hardware. “It wasn’t even that hard !“ says one researcher. Firmware available to qualified researchers who pinky swear that it will not be leaked.
Researchers set fire to laboratory with gasoline, killing seven volunteer victims, demonstrating that laboratory fires are a real risk and can carry significant consequences, especially when gasoline is involved.
this is part of the pro-active security loop. gotta demonstrate how it can break to figure out how to defend it.
our other choice is to let someone else figure it out in relative secrecy. then theyre able to cause a bunch of damage to a wide range of systems. with no defences for it. everyone would be scrambling around figuring out how to deal with it while the damage is going on. not good.
I’m totally onboard with (and an adamant user of) proactive security. But there are classes of threats that are obviously possible, and the -concept- does not need validation.
Now , a control anchored experiment with balanced and unbalanced attacker/defender LLMs, that would be instructive and useful.
The idea that an LLM can deploy other LLMs on a machine it has access to is not research. Neither is the idea that an LLM can autonomously infiltrate and expand its access over a network. I have already done both, and it’s literally just a couple of prompts and a pile of reference docs. I use LLMs to deploy LLMs on my infrastructure, and I use LLMs to analyze security vulnerabilities on my networks, including deployment of access ladders on vulnerable machines. That is SOP, not research.
If they had used a pair of identical experiments, one that was exposed to an infiltrator LLM, and the other occupied by a defensive LLM and then exposed to the same threat, that would be an actual experiment.
As it is they just threw a roadflare on a dry field, and yup,
Dry fields burn. They at least could have done it with and without recent rain.
They published only the obvious and dangerous part, none of the hypothetical or potentially useful part. Low effort, rush to publish.
I made a tiny ai bug hunting harness (<4MB) that has everything (except the model obviously). It was designed for pentesting purposes where the tiny size matters to make it more portable between environments.
The intended purpose is not to be used as a worm but it does not take a genius to figure out that with small modifications such a thing could work relatively well - especially if it uses AI keys from compromised targets. Making the agent self-modifiable is relatively straightforward task and in fact I already did that in another project.
https://github.com/chatbotkit/rook
Every Windows computer has a small rwkv model on it. Wouldn't be hard at all to get decent cpu performance from a tiny malicious harness, especially one that used the self-evolving skills features and open source models.
Malware is going to be crazy, people aren't ready for the revelation of how insecure and broken things are. Everything is held together by bubblegum, duct tape, and panicked engineers putting out fires.
In the paper they say that the worm uses either existing vulnerabilities that it has been trained on or new published vulnerabilities that it scrapes. 44% claimed success.
The paper is a bit silent on why a such a worm would need an LLM. It seems that brute forcing all known vulnerabilities, script kiddie style on each new machine is about the same.
But apparently that info is too dangerous to release ...
AV/EDRs are kinda lame but "brute force all known vulnerabilities" is definitely something they can detect.
The academic paper is here: https://arxiv.org/abs/2606.03811
It's not fully described how things work exactly, but apparently it does not transfer entire LLMs as part of the worm. Now that would be interesting :)
The abstract says:
> The worm parasitically uses compromised machines to run open-weight large language models (LLMs) to sustain its reasoning, or extend its reach for further attacks.
Thanks for pointing that out. I scanned the paper and found that in their main experiments, they use a shared GPU resource and do not copy LLMs to target machines. Apparently they did other experiments in the ablation study where they did copy LLMs.
So it's even worse than I expected. The intended worm can spread through my thermostat, and when it reaches a GPU host, it can spread even harder. Fun times ahead.
You'll just have to starve it with a bunch of thermostats that lead it towards the GPU rich honey pot where you will extract it...
I wonder if gamma ray memory corruption will induce a sort of mutation and selection effect on non-ecc-memory hosts which will make the worms effectively evolve.
I think an approach could be to use some engineered security issue or however people build botnets, and give it some AI llm that is small and minimal but comes with instructions to download models from hugging face, and some other minimal prompts and descriptions of tools. Then it could use this to grow in infected computers and try find more capable and vulnerable computers to run better capable models and also devise some minimal communication between the different points of the botnet. Perhaps set itself a goal to dominate the biggest amount of compute and have some other goal. Would be curious to see what happens.
When the worm makes someone's machine start to sound like a leaf blower, you are found out.
In the abstract, what does it mean "the attacker's marginal cost per new infection is zero"?
If you infect a machine with GPU enough to run the localLLM needed to steal another machine, you can let it burn tokens all day for free because whoever you stole the first one from will pay the electric bill.
We're getting closer to the Matrix's "We do know it was us who blackened the skies"
Ah sweet, AI-made horrors beyond my comprehension
You cannot possibly be a full-time academic and your last name be "Papernot"!
Unless your field of academia is digital. Perhaps this is why he wanted to attack printers on the network.
I'm reminded of the universal computer viruses of Steve Barnes' SF stories, which ended up infecting people too.
Doesn't Neal Stephenson's Snow Crash have a similar idea? IIRC, a computer infects human brains via language and sound.
In the 2004 Battlestar Galactica series, the explanation for why the Galactica was the only ship that survived a massive Cylon attack seems more and more likely. The ship was old and wasn't fully connected to the human's command and control systems and so the Cylon virus couldn't reach it.
ANY online device? Even assuming AI can find vulnerabilities in every operating system, there's no indication that this is actually true beyond a "here's how it could work"
This is the same nonsense that lead to article saying researchers had created a wormhole when all they had done was draw one.
I have a microcontroller with an ROM disk (i.e., physically read only). You're telling me that an AI can find a way around the physics of not being able to mutate ROM and exploit it?
I mean, if it's online it has a network/wireless card and a TCP stack along with at least some amount of RAM, so yea, in theory unless the programming is perfect it could be exploited. Now, it's not going to be used to run AI, but could very well get used in a DDOS or something like that.
Straumli blight?
Wrong zone.
sorry, but i had to do this…
is this papernot’s first paper?
Did people doubt that this was theoretically possible? Seems self-evident to me. The interesting thing will be seeing it in the real world rather than in a controlled environment where they deliberately made all devices on the network have a known vulnerability.
There’s a difference between speculation and measurement, especially since you’d have people making arguments like saying that open models aren’t powerful/fast enough to work. Demonstrating this is a useful warning to everyone (most of the industry) who’s been slacking on internal defenses because they don’t think a well-resourced attacker will target them.
Honestly with some of the denialists here a terminator could kick down their door with lazgun in hand and they'd still tell you that AI can't do that.
And for the people that think that alignment is stupid, not training your AI to think twice about writing self spreading worms is a recipe for disaster after someone gets a token stealing, resource grabbing worm going.
"Hey Honey look, I created Skynet!"
Ah yes, viral AI gain-of-function research in a secure lab. What could go wrong?
Yeah, lab leak is hard enough to contain with human viruses, but labs have well established protocols to prevent it happening.
Computing doesn't have good protocols except for air-gapping, we really just have lots of layers of best-effort detection, and billions of devices which mix data and instruction often in a careless fashion.
I used to not believe in the dangers of AI or the risk of internet-collapse from "rogue AI", but a genuine self-mutating virus could genuinely take down the internet and need an entirely new separate net. ( Or we'd discover if the current backbone actually has the power to break encryption to stop it. )
And this time, you can bet any new internet would be corporation captured. CompuServe and AOL failed because of the open internet, but we're a very different world now, governments would support the corporation led locked-down approaches for "safety".
I don't for a second believe the capability is actually there yet, but it's no longer unthinkable that such a thing could be created in a lab within a decade. Once out in the wild, there's a lot of idle compute out there to harness for self-improvement and spreading.
Next up:
Obvious pattern of using ai to replace human reasoning in a proven methodology of malware distribution, C&C, and network infiltration obviously possible, say researchers.
Researchers use AI to create the torment nexus using commodity hardware, demonstrating the very real threat that AI could enable attackers to create torment nexus nodes using commodity hardware. “It wasn’t even that hard !“ says one researcher. Firmware available to qualified researchers who pinky swear that it will not be leaked.
Researchers set fire to laboratory with gasoline, killing seven volunteer victims, demonstrating that laboratory fires are a real risk and can carry significant consequences, especially when gasoline is involved.
Just because you can, doesn’t mean you should.
this is part of the pro-active security loop. gotta demonstrate how it can break to figure out how to defend it.
our other choice is to let someone else figure it out in relative secrecy. then theyre able to cause a bunch of damage to a wide range of systems. with no defences for it. everyone would be scrambling around figuring out how to deal with it while the damage is going on. not good.
I’m totally onboard with (and an adamant user of) proactive security. But there are classes of threats that are obviously possible, and the -concept- does not need validation.
Now , a control anchored experiment with balanced and unbalanced attacker/defender LLMs, that would be instructive and useful.
The idea that an LLM can deploy other LLMs on a machine it has access to is not research. Neither is the idea that an LLM can autonomously infiltrate and expand its access over a network. I have already done both, and it’s literally just a couple of prompts and a pile of reference docs. I use LLMs to deploy LLMs on my infrastructure, and I use LLMs to analyze security vulnerabilities on my networks, including deployment of access ladders on vulnerable machines. That is SOP, not research.
If they had used a pair of identical experiments, one that was exposed to an infiltrator LLM, and the other occupied by a defensive LLM and then exposed to the same threat, that would be an actual experiment.
As it is they just threw a roadflare on a dry field, and yup, Dry fields burn. They at least could have done it with and without recent rain.
They published only the obvious and dangerous part, none of the hypothetical or potentially useful part. Low effort, rush to publish.