More like they realized how much money they were wasting letting the proles generate slop and vibe code the same CRUD app they rewrote in 5 different JavaScript frameworks a few years back.
The money is in enterprise and government. The consumer market doesn’t remotely pay enough. It’s just the same story with Microsoft purposely making Windows an unusable mess because that’s not where they make their money. It was good to establish themselves, but that market is getting dumped.
I don't think they've added enough cyber. My cyber workflow demands more trusted access for cyber so that I can use these cyber-permissive models for my cybersecurity.
It's a source of minor, but persistent, annoyance that security people have tried to abscond with the prefix cyber, morphing it into a synonym for security.
Having grown up reading cyberpunk novels about life in cyberspace, a passing interest in cybernetics (though not of the Sirius Cybernetics Corporation variety), it's frustrating to lose a 'this means computer or internet related' prefix.
Incidentally, I recently learned the origin of the term. Cyber - short for cybernetic - is from the greek κυβερνήτης (kybernetes), meaning helmsman. The original use of cybernetics is in the context of automated control systems, so steering a rudder was a good analogy. It is also the origin for the name k8s.
In my headcanon, I still read k8s as "network of cubes", as in Borg cubes, as Kubernetes itself is a poor man's Borg (as in the thing that Google runs on, named after Star Trek Borg, known for cube-shaped ships referred to as "Borg cubes"). The whole kyber thing sounds like an explanation after the fact, to detach from the Collective legacy.
I love that in the era of having LLMs summarize everything all of these companies have opted for what I call the “YouTube streamer apology video” tone and length for these announcements.
These feels more or less like a way to get in the news after Anthropic's Mythos announcement by removing some guardrails. I’m still signing up though.
It's important to keep perspective, the holes that everyone (including LLMs now) keep finding in pretty much everything are mostly the fault of running things with ambient authority, instead of using systems based on default deny, and capabilities.
I used to think we were 20 years away from a shift to Capabilities based Operating Systems, which were ----> this <---- close to being adopted widely when the PC revolution swiped them aside.
Unfortunately, I think we're about to repeat history, and we're now 20+ years out from actually solving things, AGAIN. 8(
Many (maybe even most bugs) the ais are finding are memory safety errors, which is pretty clearly not "the fault of running things with ambient authority". The data is treated as untrusted, but due to a mistake can still do something it shouldn't.
"Solve things" or actually do something useful, pick one.
If anything, maybe the security community can finally be arsed to consider ad-hoc delegation of authority as a core concept and a basic use case, because that's arguably the primary source of persistent user-level security issues in computing.
In real life, it's absolutely normal to ask random people on the fly to do something in your name, with your credentials - whether that's sending your kid with your credit card for a grocery run, asking spouse to do some bank transfers for you or set up a new computer for you, or asking a co-worker to operate some system. It's the other reason people write passwords on post-its: even without bullshit password strength rules (see xkcd://936), there's still a frequent need to share passwords with people.
Meanwhile, for the past decades, security community has been insisting on tying authority to individuals, and doing everything possible both technologically and socially to prevent authority delegation (except in top tier corporate systems, where this is technically supported, but in such convoluted, complex and broken ways that it may as well not exist - people will still resort to post-its in drawers).
Until this basic concept is recognized, I fear more broad security improvements will only result in more useful work being prevented from happening, and more people-years wasted as users figure out ways to defeat security measures so they can do their actual jobs.
I see a Security button in the what’s new box in the Codex section of the ChatGPT website. It appears to allow me to run vulnerability scans against my connected GitHub repositories.
So it seems like you just…have it once you get approved. I’m testing it now and nothing indicates I’m running a different model but it just doesn’t fight me on cybersecurity stuff
Not really. Not performant enough. Most organizations who would be interested in using a foundation model for security would either purchase the model directly or purchase a vendor who adds their special sauce or context to the model
>partner with a limited set of organizations for more cyber-permissive models.
I get where they're going with this, but still rather hilarious how they had to get a corporate speak expert pull of the mental gymnastics needed for the announcement
A 5.4 spin with slightly different guardrails is not "access to the latest models". We know this to be true from the article because they have a section entitled "Looking ahead to our upcoming model release and beyond". I wonder if they didn't just feel like they were caught out by Mythos.
This approach means only a tiny portion of the population will every qualify. Doesn't that make everyone else beholden to those few, who are beholden to OpenAI?
Another solution is to make software makers responsible and liable for the output of their products. It's long been a problem that there is little legal responsibility, but we shouldn't just accept it. If Ford makes exploding cars, they are liable. If OpenAI makes software that endangers people, it should be the same.
> Democratized access: Our goal is to make these tools as widely available as possible while preventing misuse. We design mechanisms which avoid arbitrarily deciding who gets access for legitimate use and who doesn’t. That means using clear, objective criteria and methods – such as strong KYC and identity verification – to guide who can access more advanced capabilities and automating these processes over time.
KYC isn't democratic and doesn't prevent arbitrary favoritism, it's the opposite: It's used to control people and to favor friends and exclude enemies.
> Another solution is to make software makers responsible and liable for the output of their products. It's long been a problem that there is little legal responsibility, but we shouldn't just accept it. If Ford makes exploding cars, they are liable. If OpenAI makes software that endangers people, it should be the same.
That kind of thinking is exactly why LLMs are so censored, because people think OAI should be liable if someone uses chatgpt to commit cyber crimes
How about cyber crimes are already illegal and we just punish whoever uses the new tools to commit crimes instead of holding the tool maker liable
This gets complex if LLMs enable children to commit complex crimes but that's different from just outright restricting the tool for everyone because someone might misuse it
There's always some wedge issue that means "don't punish the toolkmaker" is not politically viable. You can pick from guns to legal drugs to illegal drugs to all kinds of emotive things.
And once the wedge is in and the concept of maker responsibility is planted, it expands to people's pet issues, obviously.
The actual line of who gets punished just ends up at some equilibrium in the middle. Largely arbitrarily.
I think the classic one is pedophiles and protecting children.
If someone uses ChatGPT to create child porn or worse, to get help tracking down and meeting children, there is NO way in hell the public will accept "don't punish the toolmaker" as a principle.
"It's just a neutral tool" gets a lot harder to claim once a vendor starts specifically training and marketing the model for its ability to bypass security controls.
Yes, pentesting tools, even automated ones, are often legal. But they commonly do run up against legal restrictions and risks. They're marketed very differently from ChatGPT.
> Ultimately, we aim to make advanced defensive capabilities available to legitimate actors large and small, including those responsible for protecting critical infrastructure, public services, and the digital systems people depend on every day.
Translation: we aim to make defensive capabilities available to US and their vassals so they can protect critical infrastructure, while ensuring countries that are independent can't protect against US attacking their critical infrastructure.
Fortunately, this plan will backfire - the model capability is exaggerated and these "safeguards" don't reliably work.
First, it looks like an "AI psychosis" paper. AI psychosis has been going through armchair philosophers the way crack was going through the low income neighborhoods back in the 80s.
Second, it does not look relevant to the discussions in any way, fashion or form.
If I didn’t empirically prove it I would agree with you. Computation and semiotics are two fields that want nothing to do with each other. The SRT is through several stages of quantitative validation. For the first time in 150 years semiotics is not a philosophy. It is proven to have computational value. The SRT bolts onto any model and improves it. I’ll be sure to link you to the benchmarks when published. The relevance is it makes these treasured black boxes irrelevant.
Bold fucking claims for a "paper" that: makes an LLM with an awkward architectural tumor, and proves that it doesn't completely die on a purely synthetic task.
Further than most "AI psychosis" papers go, but still not in any way far.
And "makes these treasured black boxes irrelevant"?
With wild claims like this, either demo a generational improvement on a live model or GTFO.
I’ve been here over a decade longer than you sport. No need to bully people out when you are only 8 months in. I will be updating here when the model is live. Expect no further engagement.
Too little too late. OpenAI's shit was nearly worthless for cybersec for what, a year already?
ChatGPT 5.x just tries to deny everything remotely cybersecurity-related - to the point that it would at times rather deny vulnerabilities exist than go poke at them. Unless you get real creative with prompting and basically jailbreak it. And it was this bad BEFORE they started messing around with 5.4 access specifically.
And that was ChatGPT 5.4. A model that, by all metrics and all vibes, doesn't even have a decisive advantage over Opus 4.6 - which just does whatever the fuck you want out of the box.
What's I'm afraid the most of is that Anthropic is going to snort whatever it is that OpenAI is high on, and lock down Mythos the way OpenAI is locking down everything.
Yes. But "perform a humiliation ritual of KYC to access the actual model instead of the nerfed version of it that's so neurotic about cybersec you have to sink 400 tokens into getting it to a usable baseline" does not inspire any confidence at all.
Remember the argument that the bad guys using AI to hack systems won't be a problem because all the "good guys" will have access too and can secure their software?
reply