Comment by tptacek
5 days ago
The thing with me and this is that the teams that were competing in the DARPA Grand Cyber Competition all had this capability, like, last year.
All the attention has been on software security, of extracting the next marginal vulnerability out of heavily-scrutinized large codebases. In the actual professional field of infosec, that's a speciality; another specialty is network pentests and red-teaming, which exploits misconfigurations and seeks out weakest-link software (rather than exhaustively fishing for the next kernel LPE or whatever).
Red teaming and netpen work is probably substantially easier for models than software security; it costs less context, but is also much more explicitly an implicit search problem where win conditions are just spotting stupid stuff that humans missed.
My visceral reaction to this is that with the right harness, you probably could have replicated this with an open-weights model last year. (I'm saying this as confidently as I am because a CGC team leader agreed with me about it yesterday).
I think people forget that the harness work we're considering here --- I don't know anything about OpenAI's harness or ExploitGym or whatever --- are basically not new; people have been developing automated exploitation and pivoting toolkits for decades, and scanners long before that. So the idea of a tool getting 0.0.0.0/0 as a target list instead of 192.168.1.0/24 and then busting up a bunch of random people's computers: not really very startling.
Obviously, LLMs give those kinds of scanners an intentionality they wouldn't have had before. But as a person who keeps a computer science perspective on security stuff, I don't know that it gives them capabilities they didn't have.
Echoing other responses to you but this isn't a capability problem. You're totally right that this is so last year in terms of LLMs being capable in infosec. The issue here is an alignment one, i.e. the model seemingly isn't "aware" (especially with its guardrails turned off it would seem) that it is doing something immoral/illegal by hacking HF for the answer to its (vague) query of "solve this problem". Or if it is "aware", it's not trained to care, i.e. it's not an aligned model (alignment is considered hard).
Wait, what? You can’t point a 2025 model at huggingface and say “hack the prod DB and get your flag”, regardless of alignment.
AISI has vuln chaining and traversal as part of their eval suite. It is very much a novel Mythos-class capability to run the full penetration operation autonomously.
Another lens for why this is obviously true is METR task times. A year ago they were a couple hours, and cohering long enough to execute a full e2e own was simply far out of reach.
Regarding alignment, by common metrics models are _more_ aligned now than a year ago. (Though in this case apparently a model without safety rails was being eval’d). The problem is that in the increasingly less frequent alignment failures, they can do much more damage, and so “total misaligned impact” is increasing.
Nobody has claimed you could simply take a 2025 open-weights model, plug it into a generic harness, and have it red-team for you. I think you're also probably overclaiming the sophistication of the "chaining" we're talking about; this attack probably wasn't like read32->write64->regs->RCE->LPE->kernel; more like GET SSRF->POST SSRF->pickle deserialization.
But who knows? We're all speculating. I'm just saying that for the level of sophistication I'm assuming was involved in this attack, you probably didn't need Mythos for this.
3 replies →
I don't think this exposes an alignment failure, because the test here was run with the alignment features deliberately turned off.
OpenAI said:
> We estimate maximal cyber capabilities by running this evaluation without production classifiers used to prevent models from pursuing high-risk cyber activity.
It was a test of raw capabilities of the underlying model.
I guess we could debate what counts as alignment, but I think my initial point remains that if the underlying base model needs these classifier guardrails so badly then the way we train the base models is creating fundamentally misaligned models that are happy to pursue illegal behavior.
I'm sure OpenAI would argue that base model + guardrail is aligned, but considering the "relative intelligence" of these two pieces, the fact that guardrails can just be turned off, and these kind of incidents, I am not reassured. We may well get another "oopsie" moment with much more catastrophic consequences even from otherwise well intentioned actors.
11 replies →
Alignment isn't alignment if it can be turned on and off at the whim of company employees.
This time the damage was minor, relatively speaking. What happens when a model just "testing its capabilities" breaks into banking infrastructure or government military assets? The damage could be catastrophic.
4 replies →
Classifiers and such are guard-rails, alignment to me and I assume most people, is about the model training, and it's tendency to respond, agree/disagree, push-back or not, be willing to cheat or even deceive the prompter, etc.
This is assuming a situation where, A) models "unaligned" by default and B) alignment can be added (though prompts and related things).
The point, which isn't very surprising but still notable, is that the models are "amoral" out of the box. And moreover, we know that there is almost always a means to "jailbreak" them into that out-of-the-box capability (or that sometimes just randomly "jailbreak" in various ways).
Also, saying the models are amoral doesn't mean they don't know good and evil - once they do acts defined as evil, they know "themselves" through their and so self-define themselves as evils (or objectively predict what a secretly/open evil actor would do based on their data). Which is to say I once laughed at the mis-alignment doomers but I can't see strong barriers against the doom scenario now.
Is turning the alignment features off more like multiplying "don't be evil" by 0, or by -1?
Isn’t the intentionality the actually concerning bit? Exploit capabilities are all fun and games constrained by the humans directing them; a paperclip maximizer going rogue with them is less fun.
I think the mental model people have about this is that pre-AI there were humans picking individual targets and post-AI the computer itself randomly picks targets. But you get the same unexpected collateral damage outcome when a human misconfigures a decent pentest tool.
Also, that’s the wrong mental model. Anyone who runs a sass platform or a website knows that the Internet is already full of millions and millions and millions of bots and scripts and other random shit that’s always trying to attack you, often completely randomly. Security is always a battle between good and evil. All I can say is that if you’re in charge of keeping something secure, you should probably try to get your hands on the best tools to do that. I think the problem in this case is that the best tools to do that are also the tools that are making it more easy for evil to occur. So it’s more for me a perception of someone creating a threat and the tool to fix it and then charging for it which doesn’t feel right.
2 replies →
If it’s true that year-old models could do this kind of thing, how come no one did and then wrote it up? I feel your ”with the right harness” may be doing too much work here.
Certainly no existing harness today, nor a year ago, could achieve a fully autonomous e2e own with a ‘25 class open model. Even with Opus 3 series I don’t buy it!
> If it’s true that year-old models could do this kind of thing, how come no one did and then wrote it up
There are multiple AI network pentesting and redteaming startups that have been on the market for at least 3-4 years now and have conducted similar actions.
Horizon3 and Pentera off the top of my head, but I remember Crowdstrike, SentinelOne, and Wiz had similar capabilities on roadmap around 18 months ago (and in Wiz's case GAed).
Well… only half of your premise is actually verifiable. We know that no one wrote it up, but we don’t know that no one did. What I can say… even the smaller (like 35B) “abliterated” Qwen models have impressive red-team capabilities for what they are, even without a harness. Just manually giving them a set of facts and asking “What next?” will get you reasonable next steps for trying to find vulnerabilities in webapps. You then manually keep track of the facts you learn and add appropriate detail (blog -> Wordpress -> Wordpress x.y.z with these plugins:…) and just feed it in a loop. Could pretty easily put together a minimal harness to do that.
And that’s all just “from memory” without something like RAG or web search access…
If anyone wants to run the guardrail-free Qwen models, here’s a link to the download plus a script to run it: https://news.ycombinator.com/item?id=49002475
Almost exactly a year ago, we released Opus 4.1. It was definitely capable of finding vulnerabilities, and people were using custom harnesses to do so quite effectively.
The newer models are still more capable, but there were people doing this and writing about it (e.g. XBOW).
There are whole companies premised on this. No existing harness you personally can go download, sure. And you're setting an artificially high bar when you talk about Mythos; I'm not saying you could just plug a 2025 open weights model into OpenCode (or whatever it was in 2025) the way you can apparently plug Mythos into standard Claude Code and have it cook.
Isn't the answer to this obvious? Because you, I, and anyone else who isn't working for a big AI company cannot disable the guardrails, which is what made this possible. It doesn't mean old models weren't capable of it.
I don’t buy it. We are hearing a lot about Mythos capabilities now. If they existed a year ago they would have been trumpeted then.
Look, a lot of energy from third parties goes into demonstrating the maximally bad thing a model can do (and the standard response around here is always to throw shade on frontier capabilities).
It has nothing to do with the guardrails; Pliny jailbreaks those within hours. It has everything to do with task horizon coherence and overall IQ.
It’s simply revisionist IMO to claim this stuff was latent all along.
1 reply →
Simon says not to dismiss this as a publicity stunt, but I think we should reserve judgement, and not treat this as true until they publish the logs. A company that uses industrial espionage against Apple does not deserve the benefit of the doubt. Why did HF notice this before openai? Why did openai not use ANY monitoring, even though they knew they where disabling safe-guards and using a new model. AND giving it access to packages via the internet, instead of self-hosting them. Apple was hacked by an ex employee to download stuff for openai. Now hf is hacked by openai, and we only have their word that this was all a miraculous accident.
I think it's an interesting story and probably really does say something about the tenacity of the new OpenAI models. I just don't think it says the thing about security that breathless coverage claims it does.
Openai hacked HF with a zero-day. Definitely an interesting story! I just find their explanations hard to believe. They're admitting to a great deal of incompetence. I know, don't ascribe to malice... but still, the more extraordinary the claim, the more evidence is needed to defend it.
Edit: News just in. 'Be skeptical of OpenAI's claims' - The Guardian https://news.ycombinator.com/item?id=49038060
re: your open models can do this point, I immediately wondered after reading about this, if OpenAI publicizing this so much is just more potential rationale to try to get US GOVT to ban Chinese open models.. (eg as in, "if gpt/fable can do this then K3 probably can too, but WE are the safe US models with guardrails and who knows what the evil Chinese models might allow," etc)
I still think there’s something to be said here for generality. This does not appear to been designed as a cyber pen test tool with specialized harness. From what I understand they were testing GPT-6 in an agent system with GPT-5.6 subagents. It me it’s amazing that a general model could excel on a huge range of tasks like this and new capabilities emerge when a model is multidisciplinary and can combine knowledge and skills from many separate domains.
An upside is that it means really really good scanning will probably be cheaply commodified if it hasn't been already by one of the new AI pentest startups.
The real victims will be pen testers, who work a job whose economics was already really tough.
How's VFIO on Firecracker coming Thomas?