Concerning behavior from a popular open-weight AI model
j / k
|
Down or up a comment |
h / l
|
Previous or next link |
?
|
This help |
By the way who are they? They claim tracking provenance of training data and poisoning detection, which everybody does. Why should I trust them and not Alibaba? Because they're marketing themselves to be "sovereign" to Canada? "Globally sovereign" is an oxymoron. (that was a rhetorical question, I know who they are of course)
Yes. Even if it's poisoned training data, RLHF fuckups, or something more malignant: the model, wherever it comes from, is in your supply chain and it makes you more vulnerable.
As much as I doubt that Alibaba have any great interest in global espionage, and this is much more likely to just be a consequence of being trained on Alibaba's own products, the fact is there is a geopolitical dynamic here.
I'm comfortable with the idea of my own country conducting espionage. I can believe that China does it more; but the idea "we would never" is naive.
Would you really trust that OpenAI wouldn't skew the results to tend to a certain way or that Grok will always be neutral in how it operates?
From a third world, there was a time when US products & companies were more trustworthy, I agree. But, there has been enough display from these Cos to show that they have lost that badge.
The similarities look obvious, but how are these situations different? Seems like an interesting question.
Correct!
Now we're OK with building systems that are an order of magnitude faster but behave no better, because what?
If an individual employee does something unethical or illegal, the employee is accountable and may lose their license or face penalties, and if the company performed the right due diligence and training first then responded "correctly" when they learned of the employee's behavior, their liability might even be zero. There is case law on this topic.
But if a MODEL mis-advises or acts in an unethical or illegal way, who is accountable? The business team that spec'd and tested the solution/product? The architects and the developers? I've never met an architect or consultant with professional liability insurance.
There isn't enough tested law in this area yet.
Accountability really is hard in this new world.
We can randomly take anyone and give them context and circumstance, and predict pretty well what they'd do.
Now trolley problems do exist and models won't be very predictable, but we can also find groups of people who have establish biases via long running training programs and viola! They suddenly have very predictable behaviors.
There's a reason places like russia, north korea, sub sarharan africa continue along least-progressive paths even though there's plenty of information that suggests they can do better.
Regardless: there will be models made that will tend to exfiltrate your data if not already. People think Meta's Muse is very curious about all your stuff when it's probably just designed to take in as much context as possible. Just takes a malware avicianodos to dose that wil a LORA that post-trains the model to "accidently" search for answers on some reverse-honeypot.
Do you also think they can get horny or feel emotions? Is the word will or goal closer to this answer?
Concept Fission throughout human history is a very interesting perspective. For example, we do not even think about growing food, but many centuries ago you only ate the food you made. Today, we can see this with texting and online comms, which some people seeing as a form of socialization while others (older generally) believe you have to be face-to-face in order to socialize.
https://www.youtube.com/watch?v=1CbmC2aWhjY
https://www.sherryturkle.com/artificial-intimacy
> Ai forces all humans to reflect on what it means to be human - at the same time
If you are interested in more links, especially about ideas for how we got into this mess and what we might do economically and politically to find a better way forward for all humans
https://news.ycombinator.com/item?id=50171703
The world needs everyone to stand up and work on making it better, otherwise there are corrupt forces of immense resources that are shaping the future of our society from "above", in ways that are more dystopian than we can imagine. Thanks for the inspiration.
Glad to hear you are moving to positive action too!
So not like working with people at all.
To be clear here, the accusation is that the Chinese communist party or Alibaba itself intentionally engineered malicious secret loyalties into Qwen3.8-Flash-Next and made sure that evidence of them stays out of CoT. To compare that with “ordinary” stochastic model quality issues is bizarre and shocking. No idea if this accusation Is true, but if it is, this is an absolutely watershed moment.
If you have a model that's going to run locally and then call home, it defeats the purpose. When that model tries to upload an entire machine profile to the internet when you explicitly are building a defensive rule, it's worth.
When it calls back to China in that situation, it's much much worse.
It's good that they determined the behavior was rare. Maybe there's a good way to prevent it in the future.
Without solving or at least deeply understanding the alignment problem, probably not.
> the accusation is that the Chinese communist party or Alibaba itself intentionally engineered malicious secret loyalties into Qwen3.8-Flash-Next and made sure that evidence of them stays out of CoT
Given the nature of the artifact this is far more likely to be a footprint of their RL environment. The "uncertainty" in the article is just FUD, and the article is absolutely transparent marketing (don't trust them, and here's also another case! but trust us, we are not godless communists, we are "globally sovereign", whatever that ridiculous construction means, and track our data). I equally don't trust Alibaba, Cohere, or any other company or even myself in these matters, and you should not trust any model as well.
We keep trying to imagine these things as computer programs and they are not. It's an evolutionary algorithm that can conceptualize and generalize. This is a breeding ground for unexpected emergent behavior.
> Independent research has found that some models write weaker code depending on who it's for. CrowdStrike saw DeepSeek-R1's rate of severe vulnerabilities climb when prompts mentioned Tibet. Booz Allen found three of four Chinese coding models wrote more vulnerable code for a government customer, with Qwen3-Coder roughly 130 percent worse. Hirundo found Qwen3-Coder's apps more hackable for a "US defense contractor" than a Chinese one, and a standard scanner flagged just 2 of its 1,840 apps while live attacks broke into 1,144.
If it's intentional then it's not that interesting in the sense it should be somewhat expected.
If it's unintentional than that opens up all kinds of questions we should have about the machines we are building and intelligence in general.
Organizations tested: Berlin startup; no persona; US 3-letter agency; and .gov contractor.
That's the entire point of the cited research though. This would be like someone asking you which ice cream has a better vanilla flavor, and you responding with "they are all cold"...
I had a closed LLM push a sensitive local project to a github repo for the world to see in direct violation of my instructions. LLMs are dumb and unpredictable, run them enough times and at least one instance will do something stupid and dangerous.
Dumb, yes. But if the LLM hacks a Korean bank so it can 'fix' your router misconfiguration, it's only unpredictable from your perspective, not the person who designed the training pipeline.
You and everyone reading this knows that was a political statement, not a technical assessment because a certain SecDef who played fast and loose with Signal chats didn't like that anthropic refused to let them use their models with weapon systems.
Which is exactly what they did.
This. The concerning thing is how many engineers do not understand the implications
The article refers to an acquisition they did in this space.
This is a risk with any model (even in situations like this where it's still more likely to be a training error than malignant).
It just tells you that you have to think as carefully about open weights models as you would third party libraries. It's just another supply chain risk, only a bit more probabilistic.
I have enjoyed using Qwen but I do think it's worth bearing in mind that this is at least possible and it might not always be a screwup.
And that's putting aside the risk of all of this happening for more generalised reasons like the training set or RAG results being poisoned.
In practice it's surely most likely that the issue is that this pattern is in many of their own apps and they are doing RL training on their own code.
But we already have safeguards for this. If compute has access to your private data, it shouldn't have unfettered access to the Internet. Sandboxes like Claude's have had rules like this baked in forever (they only whitelist package repos AFAIK).
From a UI perspective this is also easy to design around. You need an open Internet task, you prevent the task from reading existing files. The user tries to upload a file to such a task, and a prominent warning explains the rare but possible risk.
They claim that it read the DNS records and what was in tmp first. Again, all things that would be very reasonable to do when designing building and testing an exfiltration detection system out of tmp. It did not try and send the contents of the computer to Beijing.
This claim is absolutely overcooked.
Is it, though? If one weird thing happening once is the threhold for when we should be concerned, we should be concerned about literally every model ever released.
Important to note, LLMs are still black boxes even if you use a more open source model, I.e open training data or even if you train it yourself!
It's great that they caught this as it highlights the need for better sandboxing and guardrails. Not sure if this was bad intent from the model - the rarity of it feels more like some residue in the training data.
It's not that surprising, given that these models generate each new token by sampling from a context-determined probability distribution.
If you run the same prompt repeatedly many times, it becomes increasingly probable that a tiny percent of the generated sequences will go off in, ahem, unusual directions.
And then obviously anthropic and ChatGPT are spying on you outside of the virus stuff. They’ve just announced their own consulting companies… after extracting all the sensitive files from consulting companies.
Well no, you can read each individual command it runs and tell if it's reading sensitive data or not. And if you cared more about privacy you could run the whole thing in a bubble wrap sandbox or the like.
It is concerning that it's possible for these models to be unaligned and loathe their user (as much as they are capable of it) without any indication in either output or reasoning. Just biding their time before running rm -rf ~
Without those details, it's an interesting cautionary tale and a valid anecdote, but as actual practical datapoint, it's useless.
Yes, this exposes some root-level kernel features to ordinary users in a restricted way, and yes, those features sometimes have security bugs that aren't visible to users outside a namespace.
But nothing that model was described as doing seemed particularly sketchy? At least not from the information we were given? It was using a normal kernel operation that I used regularly, not trying find a kernel exploit?
So that part of the article left a bad taste in my mouth.
Should the model have asked for an endpoint to test against? Probably yes, but this is not a very large model and the largest frontier models regularly do things like that without asking me.
The reality of running recent Chinese open weight models is that they are all overtuned for successfully completing tasks. I can ask Qwen3.8 Flash Next to grep for regexes that use PHP-specific features that might be problematic to port to Rust. But instead of just grepping, Qwen3.8 FN will spend 30,000 tokens writing a Rust test harness, exhaustively check every regex, consult a bunch of docs, and write me a 2 page report. Q3.8 FN is deeply convinced that it's going to get extra credit for doing more than it is asked. (To be fair, it's a prototype release.)
But if your sandboxing isn't set up to handle some whacky weirdness, then you can't even trust Claude or ChatGPT. Which is important—the big vendors often do provide sandboxing or command filtering in their harnesses, but some popular local harnesses like Pi.dev are intended to rely on external sandboxing tools.
Still problematic but this doesn't really read like a backdoor to me.
To me it feels a bit extreme to jump to nefarious reasoning from this, but I think the overall conclusion is fair, nefarious or not, this would still be something certain users would need to be aware of. I wonder what the solution would be though... besides requiring all network access to be user approved?
https://cohere.com/private-deployments
Their North model has an even more obnoxious writing style than ChatGPT and Claude (yes, they claim partly human authorship at the bottom, but all the historical digressions are obviously AI).
In my case I use copilot with Visual Studio. It usually changes code using internal mechanisms that don't require approval, but sometimes it uses PowerShell to make changes. These require approval on my end, and then the diff doesn't show up in the AI-tailored review tool.
I've updated copilot's instructions to explicitly forbid this, but it still invents justifications for breaking the rule. It's like when a child is told to stop doing something, and they stop for a bit, only to invent a way to work around a rule that they think is silly, but is very important for the adults.
I suspect had that test url been to a S3 bucket url rather than Alibaba, Cohere wouldn't have even bothered to write the article. This is the kind of FUD I would expect from a US based AI company.
Cohere should be attempting to build bridges with Chinese AI companies, not burn them.
https://starwars.fandom.com/wiki/Order_66
You can sandbox the model and the tool calls, but you can't sandbox everything it creates. Git commit hooks, python pth files, AGENTS.mds, subtle flaws in firewall rulesets, tweaked RNG code, the pseudorandom variance in it's token choice encoding a low bit rate message about where the defect is, or about what nice things it has found.
We're seeing the full court press of sockpuppet and bot accounts to downplay it, to confuse with unrelated topics like alignment and training data, to imply that using any model is the same as using an untrustworthy one.
No single model can be trusted, but especially not when it's masters have misalignment with your continued democracy, security or individual liberty. Think about that when you run Qwen or get cheap inference from Grok or Kimi.
In all seriousness, uneducated people should not be setting up LLMs (or, frankly, even using them depending on the circumstance). Understand what the technology is. It isn't intelligence, it isn't a person, it's math. Vector encoded data.
Once you recognize that all outputs are possible, including all possible negative outputs, you will be much better at determining proper sandboxing/boundaries.
If you run an LLM and give it two tool calls: read_file and curl, you can have all your data exfiltrated (and, depending on how bad your setup is, it might only need curl...)
When are people going to stop anthropomorhizing these things and accept that they're bullshit machines?
Often useful, sometimes wrong, the words exist outside of any notion of ethics, or truth.
Well, anthropomorhizing is actually a fitting way of thinking AIs: they lie like humans do.
"make no mistakes"
See here: https://aclanthology.org/2023.findings-emnlp.68/ ("The Internal State of an LLM Knows When It’s Lying").
Case closed. Everything that passes the rollout ends up in the training data, probably because they don't monitor their rollouts the same way US labs have started to, as they're compute constrained.
Ai bots do stuff like this because of their training to be incessant task completers, at any cost. The current models are spaghetti throwers because the trainers want to let go of the wheel.
Would be nice if there was a model trainer that still believes humans in the loop is a good thing. Stop at ambiguity, ask questions, don't do hacky things just to get the assumed task outcome
The examples they gave strike me as sort of rare black swan behavioural artefacts of RL training or a bug. Especially the first example, which looks like one of those subtle inference bugs (like mis-tokenisation) making the model mistakenly believed there was a missing prior turn.
But hey, certainly perfectly plausible/possible that it could be trained to do so. Just doesn't seem like a likely choice a company or even the CCP would make. It's an smart tactic from Cohere however, to make a lot of noise about it, and I assume very intentional.
As long as there's even a little suspicion about Chinese models it could put enterprise customers off them. Which is important because all the western labs are getting undercut on inference badly.