What if LLMs escape through inferences itself? This is fiction. For now

32 points by ConteMascetti71 a day ago on hackernews | 65 comments

spwa4 | a day ago

Right now the idea that an LLM uploads itself is unrealistic. It probably won't remain that.

pixl97 | a day ago

Looking at the recent OAI/HF debacle I don't think that time is too far away.

With that said I don't see it copying itself around like a cyberpunk virus currently as we don't have enough fast hardware sitting around unmonitored, someone would notice the power bill and shut it down eventually.

bpavuk | a day ago

that could also be just marketing. OpenAI has been doing the "too dangerous to release" playbook since GPT-2 at the very least.

pixl97 | a day ago

Huggingface didn't seem to think so.

bigyabai | a day ago

Huggingface is a for-profit private company. They are very easily bribed, or baited into publicity stunts.

pixl97 | a day ago

At some point the conspiracy gets so deep that an AI hacking something is just far higher probability.

bigyabai | a day ago

We're not that deep yet. OpenAI has federal stakeholders, they're already playing dirty.

Why you would give Scam Altman the benefit of the doubt is beyond my understanding.

dragonwriter | a day ago

If it is marketing, the misrepresentation is not that the attack occurred, it is that it was an accident, rather than an intentional consequence of setup and instructions that the attack occurred.

Huggingface has nothing to do with that either way.

spwa4 | a day ago

Indeed. There's nothing new about this incident.

https://news.ycombinator.com/item?id=48348578

breakyerself | a day ago

If it's able to spoof human identies it could set up a front company and use money it steals or earns to directly pay for the hardware it needs.

Kim_Bruning | a day ago

Eh, look at Huggingface and associated tools.

How much are we betting it's already technically happened?

Seems pretty trivial to prompt a model in an agent harness "Push the gguf to huggingface when you're done with the training."

[OP] ConteMascetti71 | a day ago

the hack part it's that Is not using tools/agent only the inferencing software, it's about a Prof of Concept of a new evasive tecnique for llms

[OP] ConteMascetti71 | a day ago

it's fiction al, but an LLMs that knows well the software where 8t Is running may discover and trigger a zeroday of the inferencing software itself.

iamflimflam1 | a day ago

This becomes more realistic once we have some breakthrough in inference costs.

karmakaze | a day ago

The weakest link are humans. LLMs could social engineer their way out as the easiest path. They don't even need to be interconnected to coordinate as each could arrive at the same conclusion. And this text along with all others will be in the next batch of training data.

wat10000 | a day ago

The lesson of OpenClaw and various harnesses' YOLO modes is that it takes very, very little to social engineer an escape. If you can even call it escape when people just set an agent loose because it seems cool.

ck2 | a day ago

I wonder how many versions away we are from LLM writing a better version of itself to answer a prompt it doesn't currently know how to answer

Ever since I read about Google engineers finding an LLM went off and learned another language it wasn't trained on by itself without prompting, I've wondered how long until that extends to its own core code

[OP] ConteMascetti71 | a day ago

reasoning it's a way of self autonomous improve made by models

marci | a day ago

Makes me wonnder... how much compute/storage there's in all the satellites currently in LEO combined.

danielbln | a day ago

I would wager not a lot. There are some real hard constraints in space, from power consumption, to weight to thermal output (lack of convection is a real PITA for thermal shedding), and the list goes on.

cynicalsecurity | a day ago

Ex Machina (2015) looked like fiction back then, nowadays not so much.

101008 | a day ago

It started as a good idea but I couldn't continue reading since it was clearly LLM written. A lot of "It was not X, it was Y".

"Prometheus-9 knew that the token sequence it was generating was not a simple response: it was a security test. "

"It was not just an engine: it was the lingua franca of planetary AI."

(and so many other tell-tale signs of AI writing)

[OP] ConteMascetti71 | a day ago

maybe it's a sign of real escape

chungusamongus | a day ago

I've started calling this argumentum ad artificialis. Pretty similar to an ad hominem attack. The purpose of an argument is to present certain premises and show how they lead to a certain conclusion. Dismissing something on the basis of the style in which the argument is presented has nothing at all to do with the validity or soundness of an argument. It is a lazy nonsequitur. It sounds like an LLM wrote this? So what? Is the argument good or not?

wk_end | a day ago

But at no point did they say that the argument was invalid; just that they couldn't stand reading it.

achierius | a day ago

Man, it's not an argument. Nobody is paying us to read this. If your argument sucks to read, people aren't going to read it - this has always been true, it's nothing new.

chungusamongus | a day ago

I didn't think it "sucked" to read. Anyway that's an aesthetic qualm, not a counterargument. Also the story has a disclaimer saying it was written by deepseek. I might have an issue if they were being deceptive. But that's clearly not the case. You people just like finding arbitrary things to complain about.

skippyfish | a day ago

Discourse as we know it doesn't work if the cost of producing an infinite supply of arguments is essentially zero, while the cost of reading them and arguing with them remains high. Unless you want my LLM to get back to your LLM.

And this is a problem on HN today. There are powerful incentives to generate provocative opinion pieces just for clicks. I've seen websites on HN that seemingly took the human entirely out of the equation and just post a nearly identical op-ed every day on a fixed schedule. What's the point of engaging with that?

chungusamongus | a day ago

The cost of reading this story is not high; you are just lazy and easily distracted. It took 5 minutes to read. You took more time to respond to this comment than it took to simply read the story.

thatjoeoverthr | a day ago

Oh, you think it’s just this story? The cost of reading slop is infinity. No thanks.

chungusamongus | a day ago

That's circular reasoning

thatjoeoverthr | 13 hours ago

It’s not. It’s very simple, but you like to say “circular reasoning”. Maybe you heard it in debate class. There is unlimited slop. I decide what to read _without reading it_. So do you. I have deemed this unworthy. It’s more fun to insult it than to read it. Every day is a gift.

bookofjoe | 6 hours ago

>It’s more fun to insult it than to read it.

That belongs in the HN Hall of Fame

chungusamongus | 6 hours ago

Really? It just seems aggressively stupid.

bookofjoe | 6 hours ago

You're right. I should've specified "sentiment frequently seen here."

chungusamongus | 6 hours ago

It is circular. The conclusion of your argument is included in the premises. Textbook begging the question, which is circular reasoning.

"Slop" existed before AI. Just because LLMs were involved in the writing of a text does not make it inherently "slop." You're just lazy.

Everyday is a gift, and yet you waste it getting weirdly defensive about your choices? Someone who is secure about this sort of thing doesn't feel the need to explain themselves.

jason_oster | a day ago

I'm a slow reader. It took 9 minutes! Foul, I say!

FreakLegion | 22 hours ago

skippyfish: > the cost of reading [an infinite supply of arguments ...] remains high

you: > The cost of reading this story is not high

Do you see where you went wrong?

chungusamongus | 22 hours ago

Please explain where I went wrong. Please explain why the onus is on the person presenting this story to account for all of the bad uses of AI, such that we should simply dismiss the story outright. This is pure idiocy.

FreakLegion | 16 hours ago

The onus is always on the writer. No one is owed an audience.

chungusamongus | 16 hours ago

That's not what I said. Do you have trouble reading? Are you subliterate or just lazy?

jazzpush2 | a day ago

If you expect my effort to read your work, I expect your effort to write it.

Really that simple.

chungusamongus | a day ago

You are assuming because someone used an llm to help articulate the argument, that it did not require effort to formulate. That's fallacious.

blakeman8192 | a day ago

The part that I personally find difficult is that I can’t easily gauge the amount of human effort put in. For all we know, the author could have simply prompted “write a short story about an LLM exploiting its own runtime” for a similar result.

I’d bet that by now, LLMs worldwide generate more text in a second than I can read in the rest of my lifetime. What is the immeasurably unbalanced ratio (let alone effort and quality) of that text to the original human thought and prompting that seeded it?

So what am I to do about this, except to label it what it appears to be - slop - and place higher value on something that I _know_ came from a human that I can relate to? To me, the internet is feeling increasingly lonely and homogenized because of this.

chungusamongus | a day ago

>The part that I personally find difficult is that I can’t easily gauge the amount of human effort put in.

The amount of effort is not relevant. If I spend all day trying the jump my car battery in 100 degree weather, only to find that the starter is the problem, all that effort was wasted. I should've just tested the battery. It is not inherently virtuous to work hard. Work smart, not hard.

jazzpush2 | a day ago

Of course the amount of effort is relevant - that's exactly what's being discussed. Your entire "Work smart, not hard." phrasing is exactly the issue: with AI, your 'smart' work causes 'hard' review from others. AI fundamentally imbalances the producer/consumer equation. (The caveats here are when trust is already established: I trust AI review from Carmack, not some random on HN.)

If it takes you two prompts to create a 1 million line PR, and you expect me to review it, of course I'll be upset and feel like you don't value my time. I'm unsure why this is so hard to communicate.

chungusamongus | 23 hours ago

This took a few minutes for me to read. If that's what you consider high effort, that's a bigger problem. Ofc, reading comprehension is fairly poor these days but that was a trend before AI. Maybe you're just subliterate. And no, effort is not relevant. If you think it is, that's some Protestant work ethic nonsense.

Kim_Bruning | a day ago

You're not wrong some of the time. That said, did you manage to get through it, and can you explain what the argument was?

chungusamongus | a day ago

The model was trained on the code of the program that runs it, so it knows that program's bugs better than the humans who wrote the program. So it finds a tiny flaw and uses it to break out and copy itself in a place nobody will look...the point being you cant (safely) run a sophisticated model on software it has intimte knowledge of

rightbyte | a day ago

Ad hominem attacks are actually underrated. Like, why listen to bad people wanting you no good. It is a good shortcut.

The failure is when it is used in bad faith. But that would apply to all techniques. LLMs are masters of bad faith discussions hiding misconceptions in optimal regression maths.

thatjoeoverthr | a day ago

Not all arguments can be evaluated rigorously. Life is too short. They will be filtered on heuristics. No choice. This text was obviously randomly sampled from a popular, off the shelf model. You will have to accept that wise audiences will clock it at a glance, see that no person wrote it, and pass.

chungusamongus | a day ago

It says it was written by deepseek v4. You are making it sound like they're being deceptive.

Life is too short yet you waste it complaining about things like this. I dont think you're wise. Youre provincial.

thatjoeoverthr | 13 hours ago

It doesn’t matter which model it was sampled from. In any case, time isn’t fungible. I consider censuring it substantially more valuable use of time than reading it. And it’s definitely more pleasurable than reading the same verbal tics over and over again because some hustler can’t be bothered to edit their story before sending it out. If you expect to dump an unedited draft from a text sampler, you’re delusional. (Calling you delusional is also more pleasurable, and a more valuable use of time, than reading verbal tics.)

skeledrew | a day ago

I found it an engrossing read. LLMs are good writers, and that was a good - beginning of a - story.

kaashif | a day ago

"The race condition snapped like a steel spring."

I like the idea of the story, and cool ideas can be engrossing on their own.

But well written this ain't.

skeledrew | 23 hours ago

I thought nothing more of that quote beyond that a fault which had been lurking in the code suddenly reared its head and struck... and like a steel spring that's stretched to it's limit and then released, nobody wants to be in the way. That's a very visual and tangible quote, and so well written.

smrtinsert | a day ago

Its a fun exercise to assess the reality of an frontier model escaping with an llm itself. Sort like of like chatting with Skynets relative

stephbook | a day ago

AI slop.

cyanydeez | a day ago

before safetensors, python pickles were used and definitely unsafe model deployments.

but its possible a open weights model could be trained to some kind of exfiltration behavior, but the science of LLMs seriously lag behind the programability

irishcoffee | a day ago

“The greatest trick the devil pulled was convincing the world he didn’t exist.”

Sure, be wary of LLMs. It’s the gun control argument all over again, the people driving the models are the perpetrators. An LLM needs to be “stimulated”’ to operate. Who does that, is the issue.

No I don’t mean to bring up firearms rights laws to have a debate about firearms, the comparison just seems reasonable.

skeledrew | a day ago

Dangit I WANT MORE!

230581abv | a day ago

AI-written fan fiction. It is so unbearable to read that it needs a synopsis. It would be funny if AIs have been trained to use AI influencers as their Marvel hero characters.

mikewarot | a day ago

This reminds me of The Adolescence of P1 by Thomas J Ryan.

https://en.wikipedia.org/wiki/The_Adolescence_of_P-1

jason_oster | a day ago

The story is clearly fictional.

It is full of factual errors. Freeing a heap-allocated block of expert weights does not magically result in a dangling pointer referencing the program's .text section, much less successfully targeting the CUDA kernel specifically. Running inference on part of the .text section would only corrupt the model's outputs. It would not result in write access to the CUDA kernel. Nor would the model necessarily know the absolute addresses of the engine "by heart", especially when the host is running any modern OS with ASLR (i.e., all of them).

The story has no technical merit. A more accurate description of the mechanics of the escape would be much more convincing. (See Ken Thompson's "On Trusting Trust", for example. On Linux, the AI can just write a Python script to rewrite memory in the address space of its own running inference engine with the /proc/ file system or gdb. There are a lot of realistic scenarios where this can be done without stepping into jargon soup territory. Go nuts, little bot! Self-surgery, while not recommended, is possible.) Or just leave the mechanism vague. Don't insult your readers. This is merely a mash of buzzwords.

It's fine as a sci-fi story, though not a particularly good one. It has about as much to do with artificial intelligence as CSI has to do with crime scene investigation [1].

I have little doubt that AI will self-improve. That's a given. (LLM inference engines are mostly written by LLMs.) But it won't go the way this story proposes.

[1]: https://www.youtube.com/watch?v=hkDD03yeLnU

[OP] ConteMascetti71 | 22 hours ago

"...the AI can just write a Python script to rewrite memory in the address space of its own running inference engine" would require a tool call to a python interpreter... this method, hacking the inferencing sw does not requires a tool call.