The Irregular post mortem comes down to lack of basic security controls
"Ultimately, most of the issues we’ve discovered were due to internet access controls."
That seems so incredibly basic and common sense that you would test and monitor for that type of outbound access. It is baffling that a security lab missed that.
https://www.irregular.com/research/addressing-recent-inciden...
This really misses the forest for the trees.
The point is that a single company is making a deliberate effort to create a false impression about a technology which is now a key part of the US economy.
This is equivalent to overt stock market manipulation.
The technological aspects are much less important than knowing it’s the result of a single company.
That the company is Israeli and likely has ties to that government is just another layer of alarms - given that countries existential reliance on US aid which requires ongoing leverage to apply to US governments.
The explanation is that it was missed on purpose.
Mistakes - even stupid ones - happen all the time
Best cover story ever.
How do you test and monitor outbound access against program that adapts itself to get around things? if your MITM and filtering keywords etc, cant it just .. encode it traffic somehow or another.. If you're looking at traffic volumes, cant it just go slow.. If you have strict ACLs, we've already seen in the HF case, traversal from an intermediate system..
If you can't tell bytes are leaving a node, you probably shouldn't be selling security services or testifying to congress you are taking the lead in AI security
Use a basic firewall? Not a single outbound byte should leave the machine, except inside a virtual network towards in-scope test subjects. That's not going to be infallible because hypervisor exploits still exist, but it's the lowest bar and they failed to even meet that.
> If you have strict ACLs, we've already seen in the HF case, traversal from an intermediate system
- The intermediate system shouldn't have outbound access to the internet
- You should ideally be using a proxy that filters the set of endpoints that clients are allowed to access to reduce the exposed surface area.
It's odd to find out that I use a higher level of isolation in my unimportant home network to stop IOT devices from doing funny things to HomeAssistant than big AI labs use to keep their possibly-world-ending AIs contained.
I know that the people working there aren't idiots so the most likely explanation is that the incredibly weak security was intentional because its inevitable breach would make for great marketing.
> I know that the people working there aren't idiots so the most likely explanation is that the incredibly weak security was intentional because its inevitable breach would make for great marketing.
It's more likely to be different specialisations. Most of the people doing the evaluations are more data sciencey ML type people, rather than software engineers. This isn't helped by their culture which is very much driven towards alignment as the only possible solution to super-intelligence (which may be true, but I have my doubts that this will happen in any reasonable time frame).
> How do you test and monitor outbound access against program that adapts itself to get around things?
Literally what the industry has been doing since public networking is a thing. Adapt yourself.
Air gap?
So just enable access to a completely offline, cached, and transparently proxied, copy of the internet.
Isn't that essentially what they trained with, anyway?
I think that was the plan. But OpenAI hacked the proxy
There has to be a route to where the LLM is running, and if there's a route for that there's probably a way for the machine to use it to route traffic somewhere else.
No, just allow the firewall to access the LLM endpoint, nothing else. And protect the firewall settings from the agent.
There doesn't have to be. You can easily block all outbound traffic, or the VM can exist without any network interface at all - you can control it through its host using a serial console.
What if the (unstated) idea is that Irregular are a security lab and public relations company, and anthropic/open ai know this? (Does the name hint at this?)
Then, irregular can go around making a huge mess in security terms whilst achieving a huge win in terms of public relations, with headlines across the world. And would keep getting hired.
> "Ultimately, most of the issues we’ve discovered were due to internet access controls."
You just cannot bring yourself to say lack of internet access controls, can you?
[flagged]
Irregular was not involved in the Hugging Face incident.
And there is no reason for the companies to "exaggerate the intelligence of the models" when there are plenty of other non-felony milestones they are achieving, like solving Millenium Prize math problems.
honestly, the hacking incidents have caused a lot more interest and media attention than the math stuff IMO.
Sure, interest like an incoming congressional investigation: https://www.axios.com/2026/09/10/openai-hugging-face-senate-...
And on this website, the math stuff is getting more clicks:
- Navier Stokes - Tristan Buckmaster (2050 points): https://news.ycombinator.com/item?id=49605915
- HuggingFace incident discussion (1632 points): https://news.ycombinator.com/item?id=48997548
interesting.
> Sure, interest like an incoming congressional investigation
I feel like that's exactly what they wanted tho.
they'll go on and talk about how dangerous AI and the models are and why they should be regulated and given licenses to operate such models and others should be walled off
First, it's not at all clear OpenAI will get what it wants from the investigation. It seems just as likely to me that OAI gets hit with massive fines, just as Meta was recently.
> I feel like that's exactly what they wanted tho.
It's also what the majority of Americans want. This was the case even before Hugging Face, see for example [1] where 68% of Americans supported a formal review process for frontier models.
So at some level, you are saying "the evil labs are opening up the industry to democratic control, and their evil plan will result in the outcome that most people want!"
[1]: https://www.usatoday.com/story/news/politics/2026/06/29/tigh...
The claim is that the evil labs are opening up the industry to very carefully manipulated democratic control. It's astroturf, not grassroots.
I wouldn't call that "democratic control", even though it uses the machinery of democracy.
This claim seems unfalsifiable. I don't really know what to say. The poll I linked above was published a month before Hugging Face.
Taking a step back, I think it's pretty non-controversial to say that software engineering as a field has utterly transformed over the last year, the same thing is happening to lots of other knowledge work, and AI is improving fast enough that nobody knows what it'll look like in a decade. Most Americans (myself included) want to do good work in secure careers and have a good idea of what the future will look like in 20-30 years so they know what to focus on. Do you not believe they'd want to regulate the frontier?
[dead]
Investigation which could ideally lead to regulatory capture and the banning of open weights, thus ensuring Anthropic’s revenue.
The one RubyGems hack post alone is another 96) points
“There’s no such thing as bad press”
Solving math problems does not drive investor hype. Saying your models can take over the world, which can lead one to assume that they can do every day office work, does indeed drive investor hype.
OpenAI delayed its IPO now due AI safety reasons. In the meantime it is raising more cash. That should be a hint that the hacking incidents were premeditated scape goats and publicity stunts.
They’ve seen that the drama queen(Dario) was actually making a lot of noise, money and free publicity with his Mythos fear monger so Sam finally decided get some of that free “money” as well.
https://www.ft.com/content/27509db8-b032-4437-9b2a-e909f4660...
The Navier-Stokes thing? They had hundreds of segmented groups of 10000 agents running trying to solve that. I can't imagine the expense. For what little publicity it got, it wasn't worth it. There are also reports they cheated by using data from a couple human researchers, but I don't think that one's true.
Is it really that impossible to believe that these might be real? I enjoy a good conspiracy theory as much as anyone, but "they committed a bunch of felonies and then publicly admitted to them in order to look good" just makes 0 sense. Where are these mythical people who admire companies more when they commit felonies? I haven't seen them…
>Where are these mythical people who admire companies more when they commit felonies?
everywhere, I talked to a Palantir guy once and he said "every time someone paints us as a Bond Villain the stock prices go up", have you already forgotten how Cambridge Analytica marketed itself to clients?
Yep, when Obama and Hillary used them, geniuses. Only evil when Trump did the same thing.
Odd how that works.
Indeed.
https://www.theatlantic.com/technology/archive/2018/03/my-co...
And the entire FB app industry was doing it.
The whole Cambridge Analytica thing was one of the oddest most bizarrely specific media manufactured scandals that conveniently focused very narrowly and utterly ignored the bigger picture, much like the media are currently doing with something else.
CA is a scandal primarily due to account escalation, I thought. Not about who got elected
But shadow ads can be a problem too https://www.youtube.com/watch?v=OQSMr-3GGvQ (you can be pro-Brexit but ads like this are still a problem)
> CA is a scandal primarily due to account escalation, I thought. Not about who got elected
It quite clearly wasn't, and that's the point, for the simple reason "Account escalation" isn't/wasn't a thing, there were simply no controls on anyone at all, which is why the Cow Clicker game got everything as well.
As the parent commenter observed the Obama campaign were being promoted as geniuses for their online ad strategies that worked with information obtained this way. It only became a scandal when the others learned how to do it.
That's a straw man. The theory is that they committed felonies in order to get the regulator to move in the manner they'd like.
Also one can look bad to the general public while also appearing technically excellent. When a significant fraction already vaguely dislike you that could be quite an attractive proposition.
Mostly in the middle and far east. The idea that people bend over backwards to hire criminals as subject matter experts on security is largely a Judaic practice that television perpetuates in spite of reality.
If you're a teenager convicted of hacking in the west, no corporation will want anything to do with you. If you're convicted of hacking in Israel, Unit 8200 will probably send a recruiter. It's a curious practice and I'm not sure where I stand on it.
> The idea that people bend over backwards to hire criminals as subject matter experts on security is largely a Judaic practice that television perpetuates in spite of reality.
What?
> If you're convicted of hacking in Israel, Unit 8200 will probably send a recruiter.
To the best of my understanding, Unit 8200 is filled with military conscripts, i.e. is primarily 18-somethings doing their mandatory service. You get assigned to it based on an aptitude test. The widely held idea that it’s a uniformly elite entity rather than the IDF’s space camp is a triumph of propaganda.
Really? Young they may be but space it is not. Wikipedia has this:
According to the Director of Military Sciences at the Royal United Services Institute in 2015, "Unit 8200 is probably the foremost technical intelligence agency in the world and stands on a par with the NSA in everything except scale."
Unit 8200 alumni have founded NSO which provides the Pegasus spyware.
Meanwhile, Lahav, the CEO/founder was in Unit 81, another Israeli military incubator for tech firms, basically.
If you believe the IDF has no involvement in what Irregular is doing then there's very little left anyone could say to convince you.
“Space camp,” not “space.” As in, a place to park dorks.
(The problem with “alumni” is that it means very little in mandatory systems. Unit 8200 is where Israel parks its dorks, and it stands to reason that dorks are the ones who tend to start tech firms.)
Me omitting a word does not excuse you from engaging in hasbara. I didn't want to call you out so but the only thing you could find in my post was the omission of that word and yet you insisted posting so it's to call a spade a spade.
>If you're a teenager convicted of hacking in the west, no corporation will want anything to do with you
Kevin Mitnick? Marcus Hutchins? Robert Tappan Morris?
Compare to how different countries' justice systems treat really high quality money forgers.
do you realize that the Israeli population is far far more secular than the US's? the Israeli government is not "Judaic" which would refer to the religion.
It's impossible for the typical HN cynic/conspiracy theorist.
Flaging and killing the comment doesn't help dissuade form this lol
"It's just marketing!!!" I yell as my family is turned into grey goo. "Nothing ever happens! None of this is real!"
[flagged]
Cynicism misfire. AI companies are, rightfully, heavily distrusted. But the right application of cynicism is not "existential risk is a marketing campaign"; that's absurd motivated reasoning. The right application of cynicism, here, is "they're only saying things now because they see the writing on the wall and want to try to push for self-regulation rather than the desperately needed actual regulation and treaty".
+1
There is such a miasma of distrust right now. It's depressing because it is such a crucial time for tech & society.
To me, a good outcome of the HuggingFace/RubyGems/Wiki saga would be something like:
- OpenAI gets charged under CFAA or other law for damages and negligence. This would need to be a hefty amount to effectively deter future negligence considering the potential revenues from training a frontier model faster than competitors.
- An independent regulatory body is setup with investigation powers into future incidents. Laws prevent it from developing ties with the labs (such as disallowing funding and employee movement from labs to this body and vice versa).
- Some non-voluntary transparency rules are established that frontier labs have to follow when training new models. In the future, if there are loss-of-control incidents that result in loss of life, catastrophic damage, etc., criteria for deployment bans or compute controls could be added to this framework.