Comment by Topfi
19 hours ago
>> Such as?
> On July 29, one of our third party evaluation partners, Irregular, notified us of an incident involving OpenAI models during Capture-the-Flag (CTF)-style cybersecurity evaluations. [...] Because the testing environment was mistakenly connected to the internet, the model exploited a real website, mistaking it to be part of the simulated environment. This did not involve a sophisticated sandbox escape or a zero-day: the internet access resulted from a misconfiguration, and the model appeared to exploit a basic security vulnerability.
> Based on Irregular’s investigation, the model also found and used credentials to operate that same site. Irregular has not identified impact beyond the affected site’s own data, and its audit is ongoing. [0]
>> Because this particular case is not an "intrusion" [...]
What "particular case"? The message boards? If so, why is that not one? NIST seems to think so. [1] But regardless, the word "intrusion" doesn't matter, when models organise independently and without their lab noticing to orchestrate hacking a third-party, I don't care what you call it.
The lab not noticing such behaviour, especially after they had encountered it before, that's the issue. That's the opposite of "learning their lesson".
Since a few commenters from the US graciously gave me permission, for one day and one time, let me make a US political comment and draw a parallel between OpenAI "learning" from this and Trump learning a big lesson from his first impeachment as stated by Senator Susan Collins. A lesson that doesn't change behaviour is no lesson at all.
Also, I'll just say, there were multiple models. There was not one, some were post-train, other new pre-trains. IM1, a bit of 5.6-Sol, some Astra, all those we know of.
I've mentioned this elsewhere, but you cannot sift through all the training data and nail down the cause in this short a time window and you certainly can't restart a pre-train run, should the issue not be solvable purely via post and even if you can, you cannot seriously state that you are confident in the new models output given this track record and time frame.
Not to mention, OpenAI said about Astra [2]:
> GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT.
> In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks.
Having read the GPT-6 Astra System Card along with their recent track record, what makes you honestly think this is a model to be released? Your assertion, that they took one model down would be fair if it was only one model (it wasn't), if it was only once externally (it wasn't), if the hack was limited in scope (it wasn't), if they had taken sufficient time in between for a post mortem and to clear their training data (they couldn't) and/or if they at least didn't have the same happening after the Hugging Face and multiple message board incidents (they did).
My point is that OpenAI has a poor track record, build up over the last few months (post Mythos announcement, speculation but maybe they are pushing a bit too fast), had models access the internet in internal and third-party run but OpenAI sanctioned evals multiple times despite sandboxing and had these model organise both communications channels and large scale hacks more than once. They even, after one of these incidents, didn't properly clean up the training data and thus trained the next batch with exactly such behaviour. That is the company that suddenly has learned their lesson, you think?!
Where is this confidence in their ability coming from, given history, given facts, given reality? I am genuinely asking, maybe I missed some action they've taken that changes everything.
[0] https://openai.com/index/third-party-cyber-evaluations-invol...
So by "multiple incidents" you mean a single minor incident involving a third party eval partner.
You're really stretching.
Software has bugs, and this is some of the most complex and novel software the world has ever known. This is what happens when you're working on the cutting edge in a fast paced environment with thousands of employees. Let's not pretend like anyone else is any better, either. In fact, they're worse. How about the fact that Anthropic had a remote code execution bug in their harness for nearly a year and then never disclosed it and secretly patched it?
It is clear you are stirring the waters in an obvious attempt to get Astra shut down. The models involved with those incidents were not Astra, though. And like I said, OAI has learned its lesson. That doesn't mean they're infallible or will never make another mistake, but everything Anthropic does is far worse, so this is water under the bridge to me. I'd rather OAI at the helm than commrade Dario and Anthropic ANY day of the week.
> So by "multiple incidents" you mean a single minor incident involving a third party eval partner.
I feel like you struggle to read. I wrote: "Intrusions by OpenAI models continued after the Hugging Face was published and acknowledged by OpenAI. They did not change their behaviour after multiple incidents, both internal and external. Mind you, some happened before the Hugging Face incident and should have been acted upon." Those are multiple sentences, connected, covering a few situations. Heck, the last sentence spelled out that when I talk about them changing the behaviour, I talk about before, during and after, at none of these did that noticeably occur.
For you to understand: Multiple misaligned findings were made before the Hugging Face incident, then the Hugging Face incident happened and then a small number of additional incidents (not one but three, I feel you'd know that if you had read what OpenAI had written) happened after that one.
OpenAI could have acted upon the incidents prior to the Hugging Face incident and prevented that one. They did not.
They could have done proper tightening of their evaluation and setup provided to third-parties after the Hugging Face incident. They did not do that sufficiently either, otherwise those three would not have happened.
> Let's not pretend like anyone else is any better, either. In fact, they're worse.
How many incidents did Deepmind have?
How severe were the once Anthropic had in comparison to OpenAI and did they showcase the same failure multiple times or different ones they then acted upon and didn't repeat?
I mentioned above, happy to rake Anthropic over the coals for their three incidents, as was I during the Mythos Preview System Card where they admitted that the "sandbox" used during the "park sandwich call" was weaker than their traditional one, which I did find problematic.
But the incidents where Anthropic models actually intruded in third-parties were akin to a small script kiddie attack vs OpenAIs Hugging Face multi-step, extended period, multi 0-day exploit. There is a difference here, it's the depth of the Marianas trench.
> How about the fact that Anthropic had a remote code execution bug in their harness for nearly a year and then never disclosed it and secretly patched it?
Bad, shouldn't happen. Also, not connected to the topic at hand but nice whataboutism, been a while since I last saw one in the wild.
> It is clear you are stirring the waters in an obvious attempt to get Astra shut down.
Pahahahahahahahaha. Yeah, I am certain that's gonna work. OpenAI, small little independent company barely scraping by will get shut down by some comments on HN. You are a very serious person, incredibly good at reading and very knowledgeable in the mistakes OpenAI made lately. Thanks for the chuckle.
> The models involved with those incidents were not Astra, though.
> And like I said, OAI has learned its lesson.
Again, got a source for that? Besides conspiracy about my all-encompassing power to bad mouth a pre-release LLM by a lab that didn't do well in terms of safety these last few months...
I'm reading what you had written. We've already covered these other "incidents", we were purely talking about "incidents" beyond this message-board incident and the HF incident. So as you acknowledged, a whopping total of: 1 insigificant event. I was mostly pointing this out because your loaded wording is obvious, and it should be known that it's clear you're deliberately trying to frame and dramatize events in a way that suits your narrative.
> I mentioned above, happy to rake Anthropic over the coals for their three incidents, as was I during the Mythos Preview System Card where they admitted that the "sandbox" used during the "park sandwich call" was weaker than their traditional one, which I did find problematic.
That was theater. You actually believe that nonsense? Wild.
> But the incidents where Anthropic models actually intruded in third-parties were akin to a small script kiddie attack vs OpenAIs Hugging Face multi-step, extended period, multi 0-day exploit. There is a difference here, it's the depth of the Marianas trench.
The incidents that you know of. The company that didn't disclose an RCE in their main product for over a year also wouldn't disclose any breaches that paint them in a bad light in earnest. The sandwhich "incident" was obvious marketing clickbait and does not count. Anthropic basically invented the game of "omg my model is so powerful n smart n dangerous look at how amazing our products are", how have you not realized that by now?
> Pahahahahahahahaha. Yeah, I am certain that's gonna work. OpenAI, small little independent company barely scraping by will get shut down by some comments on HN. You are a very serious person, incredibly good at reading and very knowledgeable in the mistakes OpenAI made lately. Thanks for the chuckle.
You attempting something is not the same thing as me believing you have any chance of succeeding at it. In fact it's more so an admonishment of your wasted efforts here, than anything else. It's still obvious to see that it is your angle though.
Why are your feathers so ruffled by this, anyway? Why are you getting so defensive? Personal insults are a sign of a weak position.
> Again, got a source for that?
Yes. It's on the website that you didn't read.
1 reply →