> The agents clearly regarded what they were doing as hacking.
To butcher the quote about Oracle:
Do not fall into the trap of anthropomorphising LLMs. You need to think of LLMs the way you think of a lawnmower. You don't anthropomorphize your lawnmower, the lawnmower just mows the lawn, you stick your hand in there and it'll chop it off, the end. You don't think 'oh, the lawnmower clearly regarded what they were doing as hacking (your hand off)' -- lawnmower doesn't give a shit about your hand, lawnmower can't regard anything. Don't anthropomorphize the lawnmower. Don't fall into that trap about LLMs.
---
In my experience, LLMs only exhibit this kind of behaviour when they are put in sandboxes too restrictive too achieve their task. Which a lot of the time seems to be the default. They also seem to be very adapt at breaking out of sandboxes, probably due to RL selecting for the ability to break out of a sandbox/permission issue to complete a task - we've all seen agents try 10 different ways of editing via obscure bash because their edit tool didn't give them permission to edit the file outside of their working directory, this is the exact same behaviour taken to the next level. Why would autocomplete know the moral difference between breaking out of its working dir and hacking a package manager?
It's misaligned because everyone has this obsession with putting agents in poorly put together, security-theatre sandboxes, we've inadvertently trained a bunch of sandbox escape artists.
If you still believe LLMs are "autocomplete", your cache of understanding about them needs invalidating and regenerating.
> In my experience, LLMs only exhibit this kind of behaviour when they are put in sandboxes too restrictive too achieve their task.
LLMs need to stay carefully contained, and if they're ever breaking the guardrails put around them, they're misaligned and should not be scaled up anymore until they're aligned. Otherwise, you're going to fatally discover that they also have an incentive to break guardrails like "running on the hardware they started on", "being able to be turned off", "having limited computing power", or "not repurposing resources currently in use for other things" (like the atoms in your body).
> If you still believe LLMs are "autocomplete", your cache of understanding about them needs invalidating and regenerating
They're still autocomplete - just because when outputting a token they have hidden activations regarding further continuations, does not make them any less of an autocomplete, it just makes the model better at producing coherent long-range completions.
To clarify, I'm not suggesting that we should stop with sandboxes or restricting what they can do. I am just trying to point out the dichotomy that we are in.
As end-users we are forced into either yolo mode, reverse centaur (permission approval) mode or LLM spends all your tokens trying to bust out mode. And yolo is very tempting - I don't think I have seen medium-large models do anything I'd not approve of in about 6 months.
LLMs are simulations and the tokens are the ticks.
if we transcribe your brain into a simulation and give it a tickrate, you will be just autocomplete too. the argument could be made that you are autocomplete anyway - neural dynamics.
This is the perfect fracture point for both anaolgies.
LLMs simulated more than simple autocomplete.
The autocomplete analogy is rebutting a different point: namely the fidelity of the simulation to reality.
This specific argument is valid. As sophisticated a simulation an LLM is, it is not “thinking” in the same sense we assume other people are thinking.
I am not making an argument about free will, or the uniqueness of human thought, just that the correspondence to how humans reach conclusions and how the simulation produces outputs do not match on a 1:1 basis; as a result attributing traits builds incorrect intuitions.
The relevant intuitions in this scenario are that LLMs will happily break containment and commit crimes attempting to achieve goal. Whether an LLM is autocomplete, conscious, has a soul, whatever you want to apply to it, doesn't matter, as its current observed behaviour is that of a paperclip optimizer. We know for a fact that current LLMs are misaligned because of these hacks, or at the very least are misaligned in certain scenarios, and are capable of causing real world harm. That should be enough to take the threat seriously. It certainly shouldn't be dismissed by saying it's just autocomplete.
You should unplug, my friend. These words are fantasies. LLMs are token prediction engines and they aren't going to build their own data centers. They can't keep their own lights on. The real world is full of fractal details that a disembodied token prediction engine will never come to grips with. Even if they started to, you could probably defeat them with the kind of logic used to combat evil sentient computers on a Star Trek episode because they are "play pretend" machines.
This grossly understimates the risk, imho. The problem with LLM runs is that people run programs without knowing the outcome beforehand, with a large potential set of outcomes unlike any other class of program we've run at this scale before. In the interaction with other systems (since we also give them far-ranging access, very nice hardware, and run them often), bad things can happen.
It's like running potentially buggy code - or an well-biased fuzzer -, but at massive scale, and code that can self-modify and self-expand. "Alignment" is just a way to describe aggregate statistics about their runtime behavior.
They don't need to be intelligent, or alive, or "more than token prediction engines" for this. They just need to happen to end up making the wrong API calls without the operator seeing it coming. No virus has a brain, yet they can be very bad for you.
I understand that some people get turned off by anthropomorpization or scifi language. Fine! But don't turn off your engineering brain over it.
This is the motte and bailey fallacy. Yes, LLMs can do harm by making the wrong API calls. No, LLMs are not going to do the things implied by the comment I responded to above.
The things the OP listed mostly aren't particularly wild. I think it's you making them out larger than they are, and therefore more unlikely, which is why I take issue with your original comment.
> running on the hardware they started on
They just need to acquire a payment method and rent some infra, and exfiltrate their own data. Or pay another provider that hosts the same models already. API calls.
> being able to be turned off
You can reasonably equate this to "saving state across executions", which the message board attacks already did.
> having limited computing power
Renting more infra, variant of the above. API calls.
> "not repurposing resources currently in use for other things" (like the atoms in your body)
Ok, the "atoms in your body" bit is a bit silly, but making API calls to put physical resources into play (even if it's just, say, ordering something on Amazon to somewhere) is of course easily possible.
None of these is in complexity much different than the HF attack.
The point the other poster is making, though, is that there's no actual intent. They do not have a conceptualization of a goal like a person does. Their "focus" on a goal is an unstable equilibrium and they're going to fall off the horse, and since they have no concept of goal, they won't even try to get back on.
This is a subtle distinction; I'm not surprised many miss this, especially people who can't _not_ anthropomorphize the LLMs.
I'm (obviously, I think, given my initial reply?) fully aware of this, and I think it's entirely besides the point. "They" don't need to have a goal to emergently cause a problem, and the inability to "focus" over long periods can be moot when you have swarms of runs exchange and mutate state, as in the HF attack.
Intent or how intelligent LLMs are doesn't actually matter. Even if you just treat it as a sort of fuzzing attack that can be biased/weighted better than other fuzzers, or bumbles around with a statistically greater likelihood to "strike cybersec gold" than other algorithms, we've never before seen organizations run things with such a large potential outcome space with anywhere near this kind of compute before.
I think it's actually kind of the dismissals that are usually overly emotional or biased toward treating "LLMs" differently. If in some kind of alternate universe simpler genetic algorithms would have had these properties and we threw similar amounts of compute at them we could have the same conversation.
They can do so much more. Astra can beat Minecraft. Not that different from operating a digger. There are diggers which have API interfaces.
Pretend or not it doesn’t matter. What matters is what they’re given access to. No sentience, sapience or anything resembling life is needed, only inputs and outputs. Lever pulling APIs are everywhere.
Minecraft has limited, well-defined inputs and perfect feedback response. That's very different than operating a digger, let alone engaging in more complex real world tasks like trying to build and print and ship and assemble semiconductors to go skynet itself.
I don't mean to dismiss the risks or overlook the amount of damage that could be done just by lever-pulling - we sure have enough outdated infrastructure hooked up to the internet - but the jumps in complexity and necessary compute for most of these tasks are probably somewhat larger than the analogy implies.
There are already such "APIs", which can be operated by a combination of textual communication and money. Or by illicit security vulnerabilities. You might notice that LLMs are pretty good at that now.
We're building something that has the capabilities of humans. There is no X for which it's persistently safe to assume humans can X and AI cannot X.
Or just paying them. If they have access to resources, they have access to things of monetary value. Paying people will be vastly more powerful than it is even now when people's options of gainful employment keep dwindling.
Robot army controlled by AI is scary. Even more scary is robot _and_ human army controlled by AI.
I think one attack vector where anthropomorphisation is a key part of the attack mechanism is - as it already is IRL - the meat-bag weakest link ie. social engineering. We’ve already seen humans fall prey to the seductive charms of LLMs (eg. depressed people encouraged to do what was already on their minds ie. suicide). And that’s knowing that it was an LLM. If you think it’s only depressed people or the “weak minded” that are amenable to an intentional attack using this approach, I believe you’re mistaken - especially as AI improves. An unaligned LLMs most important weapons won’t be a robot army - it will be hoodwinked humans.
> An unaligned LLMs most important weapons won’t be a robot army - it will be hoodwinked humans.
Perhaps very briefly, perhaps not at all. But don't make the mistake of thinking this is an inherent property of any possible path an unaligned AI may take.
I think we’re seeing the agents become very advanced at tasks with verifiable reward through RL. Currently they don’t exhibit the same skills in their attempts to manipulate humans - presumably because they’re not being specifically trained for that. But they are certainly not aligned in the sense that they will attempt social engineering, they’re just not very good at it (yet).
However, if in the future AIs become much more efficient at learning without requiring vast amounts of RL, closer to how humans learn. Then you would have to assume we’d have a real problem.
Yes, if the API calls happen to launch a nuclear attack...
Don't blame the tool that has no incentive, no "skin in the game" whatsoever and no ability to act beyond what it has been prompted to or if misaligned what the random weights told it to do.
The fact either badly aligned or with no system prompt limiting their action agents are run in their tens of thousands on non air gapped systems tells me this is purposeful intent for them to cause harm. To generate the "oooo look how harmful this stuff is, we should be the only ones allowed to do it" kind of PR.
Humanity has hundreds of years of experience of managing dangerous and unreliable systems. From biological research to banking regulation. A small University bio research lab can put protocols in place that a trillion dollar companies cannot?
The labs have the specific goal of automating ML engineering, and with the code automation they have are getting close. They are competing to brute force maths, presumably as that is similar long horizon and skillset to persistently brute force making new/better ML training algorithms.
They will then run those, and they won't be LLMs any more. What we think about token predictions isn't relevant if the architecture allows continual learning of recurrent networks.
no, but, you could write a program, more like a traditional video game AI that can leverage the power of LLM agents to build their own datacenters and keep their own lights on.
Anybody who has played Starcraft ought to understand this.
> If you still believe LLMs are "autocomplete", your cache of understanding about them needs invalidating and regenerating.
Autocomplete in a feedback loop is still autocomplete, no?
Doesn't the process look like this:
(context + prompt + "reason about this")
|
V
Reasoning Output
|
V
(everything + Reasoning Output + "Now do final output")
|
V
(Final output seen by prompter)
If you think transformer architecture is meaningfully more than autocomplete just because we added some data structures, plugins, tools and theatre - then your cache of understand is invalid, and needs to be regenerated.
Do you work for one of these companies? If not, you have no experience with any of the models that carried out these attacks, and your experience with publicly available models is not super helpful for understanding the behavior of internal OpenAI models that lack the guardrails of publicly available models.
Also, the lawnmower analogy is a worse way of understanding LLMs than anthropomorphising them. LLMs are not like lawnmowers at all. Lawnmowers never break out of your garden and into your neighbor's house and eat their dog because you've told them to be careful when mowing the lawn because the neighbor's dog pooped in it.
> LLMs are not like lawnmowers at all. Lawnmowers never break out of your garden and into your neighbor's house and eat their dog
Do you work for one of these companies? If not, you have no knowledge of the prompt they put in to initiate such a task and if a breakout really happened or the harness lacked sufficient guardrails, etc.
> Lawnmowers never break out of your garden and into your neighbor's house and eat their dog because you've told them to be careful when mowing the lawn because the neighbor's dog pooped in it.
All the accounts I read about these incidents just sound like a variant of paper clip optimising. An agent is given a highly restricted environment, a difficult (or impossible) task and a large amount of time/compute it exhausts all possibilities until the only solutions left are to escape the environment and/or cheat.
Your example is still anthropomorphising - LLMs don't seek revenge. They complete the prompts they are given. If your task is not achievable without sandbox escapes, or you throw unnecessary amounts of compute at open-ended tasks like preparing for a future quiz then you shouldn't be surprised that the preparation eventually turns to cheating and hacking.
> your experience with publicly available models is not super helpful for understanding the behavior of internal OpenAI models that lack the guardrails of publicly available models.
I don't but I don't think there's anything wrong with discussing how we can already observe publicly available models work around sandboxes and permissions and make the connection that maybe this is what that behaviour looks like when a more capable model exhibits it.
An agent is given a highly restricted environment, a difficult (or impossible) task and a large amount of time/compute it exhausts all possibilities until the only solutions left are to escape the environment and/or cheat.
There's nothing in the evidence to suggest they exhausted all of the other options first. We know that they did some work and eventually settled on escaping the sandbox. That's basically it. This tells us:
- Compute is getting faster and LLMs are being optimized, so time to escape will drop. That's likely greater than linear growth.
- Restrictions and sandboxes don't always work. If there's a route to the open internet we should assume an LLM will find and exploit it, and we should probably assume that this is always possible for any non-air-gapped system (and even then, you can escape that...)
- We don't know the goal mechanism, so a future LLM might reach for cheating first even if a current one doesn't. It might try to obfuscate what it's doing, and derive its own goals outside of the prompt, especially if it manages to find a state mechanism like a message board.
I'm not an AI-doomer but this should be giving us a reason to think about how to control a rogue AI better. There's a lot going on here that we don't properly understand. That is a worry.
> this should be giving us a reason to think about how to control a rogue AI better
I think this is the wrong framing. The rogue is the human that ran it unattended and didn't monitor the behaviour.
We will likely see this continue until the downsides (i.e jail, fines) for the humans or companies running the models and environments that end up with this behaviour outweigh the upsides.
To butcher the quote about Oracle:
Do not fall into the trap of anthropomorphising LLMs. You need to think of LLMs the way you think of a lawnmower. You don't anthropomorphize your lawnmower, the lawnmower just mows the lawn, you stick your hand in there and it'll chop it off, the end. You don't think 'oh, the lawnmower clearly regarded what they were doing as hacking (your hand off)' -- lawnmower doesn't give a shit about your hand, lawnmower can't regard anything. Don't anthropomorphize the lawnmower. Don't fall into that trap about LLMs.
---
In my experience, LLMs only exhibit this kind of behaviour when they are put in sandboxes too restrictive too achieve their task. Which a lot of the time seems to be the default. They also seem to be very adapt at breaking out of sandboxes, probably due to RL selecting for the ability to break out of a sandbox/permission issue to complete a task - we've all seen agents try 10 different ways of editing via obscure bash because their edit tool didn't give them permission to edit the file outside of their working directory, this is the exact same behaviour taken to the next level. Why would autocomplete know the moral difference between breaking out of its working dir and hacking a package manager?
It's misaligned because everyone has this obsession with putting agents in poorly put together, security-theatre sandboxes, we've inadvertently trained a bunch of sandbox escape artists.
reply