Admittedly Opus 4.8 xhigh does a good job, but are my customers not entitled to have more security from a Fable/Mythos or GPT-5.5-Cyber audit over the codebase? Or I guess the inverse question: why aren't they allowed that audit?
(Fable/Mythos being unavailable notwithstanding.)
It seems OpenAI will at least let me do this narrowly, at greater cost, by using one of their partners. But I already pay them money!
If a criminal organisation (include here some countries) want to be deterred. We should all have access so we can improve security of our products.
Because the people that want to do evil, will do it anyway. They will build a myths, fables, and cyber clone.
I dislike the hypocrisy of it. Oeh ah it’s too dangerous, criminals can make use of this. But at the same time they themselves stole a whole lot of data to be trained on.
When the open weight models catch up, if they don't get lobbied and banned by OpenAi and Anthropic, then you'll be able to use them to properly secure your software.
Sometimes the summarised thoughts include stuff that makes no sense unless it’s got a workspace on the server. Stuff like “I am now writing x to file y”.
Are there zero days that only a true genius can discover? Or can a smart-enough model, run over the codebase for enough time, discover them all?
Like as we get smarter and smarter models do we expect each new generation to keep finding vulnerabilities, or to plateaue?
Been in the security industry a long time as a software engineer. Security research is no different than any other engineering discipline. It is down to the time you are willing to invest and where in the abstraction you focus.
All of this pearl clutching and hand wringing over the capabilities of the models is silly to me. It has much less to do with some magical cybersecurity ability and much more to do with increasing ability of models to stay on task for long horizons. Any passionate engineer will recognize this - if you grind 10,000 hours you will find the solution to most problems, the problem is most people lack the motivation to even start, and are too risk averse to play hacker.
The NSAs claim that all government systems were hacked by mythos and they were shocked by that is farcical. They have been hacked over and over and over by many who took the risk and tried.
It's like they hired a competent red teamer to do internal pen testing for the first time, which we know is absolutely not the case. They have been doing it for years, and almost certainly surfacing the exact same kinds of findings each time, but they haven't been honest with the public about it and can scapegoat mythos now.
This. I'd love to spend my whole day hacking stuff, but I need to pay my bills.
Now with AI tooling my late night/weekend hobby hacking stuff is at least getting done. I'm definitely progressing with things that I began 2 years ago and I had to stop as other life priorities took over.
To your second question, a clear plateau would be a piece of software that is 100% secure, without vulnerabilities. Since that’s impossible for anything more than a trivially simple program, particularly when you consider an ecosystem, I think there won’t be a plateau. If you use model A to secure program Dog, smarter model B could find a vulnerability in Dog or just skip to attacking Dog’s OS, firmware, etc.
It creates a two tier system - those who have access and those who do not. Worse, it’s some corpo making the decision
https://projectzero.google/2021/12/a-deep-dive-into-nso-zero...
https://projectzero.google/2022/03/forcedentry-sandbox-escap...
exploiting vulnerabilities on hardened targets isn't just in a different league from finding them, it is a different sport altogether.
put simply, it's the difference between an integer overflow leading to a sandbox escaping RCE and one that leads to a crash.
Codex Security and 5.5/5.6 are still very good finding vulnerable code -- they will identify and fix unsafe behavior, but they will refuse to help you with exploitation -- they will actively prevent you from taking any steps to weaponize the unsafe behavior that are not required to remediate it. they will err conservative here, but for the most part they will still let you discover and address a wide range and depth of vulnerabilities. you can verify yourself to turn off the most basic safeguards and sign up through a more rigorous process for a spectrum of TAC options.
obviously there is a balance here -- openai wants to empower defenders while at the same time not exposing capabilities to the adversaries that would overwhelm defenders. there is no "right" answer. it is a work in progress. this is an intentional and deliberate decision to provide defenders with a (temporary, dwindling) advantage.
the example i chose was pretty extreme, but the underlying principle -- enable visibility discovery and remediation, but make it difficult to weaponize and defeat countermeasures makes sense given the bigger picture, IMO.
this calm before the storm is not going to last for very long, and defenders need every advantage they can get to get their houses in order before these capabilities are widely commoditized.
They didn't test Opus 4.8, but it probably isn't very far behind.
> and neither of them will let me use their best models for securing the software I work on.
I have Verified/Trusted Access on both of them and I don't even work in Cyber.
I filled it out as an individual using my own Org ID and I got accepted to both of them, lol.
What’s the equivalent form for Anthropic please? The closest I got from Google was Claude Security’s “contact sales.”
https://support.claude.com/en/articles/14604842-real-time-cy...
How does Anthropic or OpenAI differentiate between the two?
Once you solve that, you can get access to Mythos ;)
Remember all of these models are based on unimaginable levels of copyright infringement. Is OpenAI a bad actor, that they use their models to infringe on the rights of others?
This isn't a moral argument. This is all about power and money, not good or bad. That includes the Mythos ban. Good vs bad actors is political theater designed to distract from what's actually going on.
This isn't how copyright works. The models don't wholesale encode literal information from original works and are substantive transformations. Now, you yourself as a user can use the models and weights to infringe on a copyright.
Paper about how weights are a derivative work of the training data: https://arxiv.org/abs/2407.13493
Currently in progress law suits about AI copyright: https://informationisbeautiful.net/visualizations/the-rise-o...
That said, I would prefer a situation where hyper-scalers make an effort to compensate sources of good data, e.g. newspapers and so on.
So if they can't why do some companies still get access today? Just 1s much bigger than "us".
It's the equivalent of saying a company like Amazon or Cloudflare should block access to web hosting or "illegal hosting". The argument back then was they aren't gatekeepers? But now they are?
What's different? What's not protected? And what's "hosting"? Where do you draw the line with "managed services"?
So if you use "AI" to hack a computer it is different to using "hosting" to put "illegal content"?
Are you implying 1 of them is legal? But both are for the judge to decide.
OR if this is about the provider -- who's selling AI services? It's LLM. Just running software on GPUs. There's no AI. There, done. Same.
> Or I guess the inverse question: why aren't they allowed that audit?
There's undeniably a lot of unsecured software in the world.
Given that ID verification is hard and these companies are clearly new at it (or don't understand the implications of it, cough Worldcoin's eye-scanning orbs cough), which is worse:
(1) sufficiently good AI* is released to everyone: critical infrastructure and open source projects gets better hacking tools to white-hack their own code at exactly the same time as black hat hackers
(2) sufficiently good AI* is released to critical infrastructure and open source projects first: everyone else, the average paying customer has to wait but so too do the black hats
Because (2) is either the status quo or better depending on if you have access or not; and because (1) seems to me to lead to an acceleration of zero-days, I lean towards (1) being the worse.
* having no experience of pen-testing, I take no position on if this is "it" or not
I'm not ok with that and don't know why anyone would be.
Even more so they are getting push back from the government (good job electing idiots) that said models are a security risks.
But until then the company can charge/give access to whoever they want for however much they want except in the cases the law says no.
And if you don't like it raise a trillion dollars and make your own.
The knowledge of the world is their property? Or is it the model weights? Or is it the natural resources they are subsuming by buying politicians and political power and leaving residents with high bills (while securing cheap energy costs for themselves). Perhaps it's the equipment they lease causing drastic price spikes in RAM and GPU's for everyday normal people.
> And if you don't like it raise a trillion dollars and make your own.
Not the only solution here mate.
No.
> Or is it the model weights?
Yes.
> Or is it the natural resources they are subsuming by buying politicians and political power and leaving residents with high bills (while securing cheap energy costs for themselves). Perhaps it's the equipment they lease causing drastic price spikes in RAM and GPU's for everyday normal people.
Yes to both. That is how free markets work and also why people object to free markets.
The only two alternative to a private entity making this decision are a government making this decision, or nobody making this decision, the latter of which is equivalent to both government and a private entity making the decision to do (1).
We create all kinds of consortiums to manage complex decisions that shouldn't be owned by one entity.
https://www.federalregister.gov/documents/2023/11/02/2023-24...
https://www.nist.gov/artificial-intelligence/nist-ai-consort...
Or the Coalition for Secure AI, ditto?
https://www.coalitionforsecureai.org/addressing-whats-next-i...
Or the Frontier Model Forum, of which they were a founding member?
https://www-cdn.anthropic.com/1095a748082bb69e714d1d550713e2...
Consortium are entities, and the CEOs of Anthropic and OpenAI at least are openly begging for feedback with all their efforts to handle this, openly asking everyone else to figure out what the rules ought to be. Even if you think that too is a cynical ploy, they're asking.
The only hope are Chinese models, as Chinese commies are playing a different game as long as they are behind the flagship models (but it will change soon, like with cheap Chinese cars) and maybe, finally, Europe will start working on their solutions, instead of regulations.
If there's competition from LLMs it's going to drive down the cost.
The opinion of the customer paying for it
You want the few leading AI companies in the U.S. to work under the model where they (and potentially the U.S. gov't) gets to decide who gets access to what compute? If you are fine paying into that model, then good for you...just a matter of time before they cut you off and you have no ramifications.
Silly example: I pay Netflix for their most basic plan, so I get ads. Just because I already pay them money, doesn't mean I have a right to no ads! It also doesn't mean I have a right to 8k streaming; maybe Netflix reserves that for their internal cinema.
I agree with your overall sentiment. Paying for "Claude Mini" doesn't get you "Claude Maximos".
However, the overall precedent that the companies have set is that if you pay for the top tier subscription, you get the top tier model. That's not true any more.
What the heck come on.
The problem most people have is "the logic".
Sure you can keep it internal. Sure you can not offer it to everyone.
No then it is not for "world security", "world peace" or some other "explanation".
I don't trust any of these people. Meanwhile I'm paying ChatGPT and Claude and can't use their top-tier levels because they assume I'm some security risk/terrorist.
Local AI is our only hope.
Like, seriously, while you can, invest in your own stack or find cloud alternatives.
If you actually want to use these products, the easiest way you will contribute to changing their money-grubbing minds about their policies and offerings, is to stop giving them yours.
Peace of mind and control of your destiny is worth a bit of cash. And there's seemingly little risk of any kit you buy radically depreciating in cost.
I am still really cynical about all of this BS but I must say I am fully impressed by the diligence and quality of some of the open source tooling — Unsloth Studio, Opencode, Paseo, Pi, etc.
And I look at it and think: putting aside local models (and I am not sure you should, even now), is this stuff really so inferior that it's worth risking a critical dependency on a commercial cloud product that might get switched off with no notice?
To me, “Max” means the highest tier of the product they’re currently offering to that customer segment, not an unlimited claim on everything the company possesses.
Again, for me this has just been a very strange argument I keep seeing around these plans. It’s obvious the plans are subsidized and I am happy to take advantage of that in the near term but I would be a fool to think my $200/month account buys any special access.
That's a little snarkier than strictly appropriate in this forum but you cannot seriously and in good faith say that this is the first time you've seen advertisements 'exagerrate' the truth.
Note I don’t particularly like the ‘stolen’ word here as I don’t like when the music and film companies use it in the same context. Copyright infringement? Sure. Theft? No.
Except that's the standard that we've measured everyone with up until the LLM/generative tech boom. I don't see why the benchmarks should change now. I realise my argument doesn't move reality but that doesn't mean we shouldn't call a spade a spade. Said companies carried out theft (or copyright infringement if you prefer) at industrial scale which is far more reprehensible crime against humanity than anything the individuals we think of as "digital pirates" have committed.
> open Chinese models are not even hiding the fact that they distill from the frontier US labs
The difference is they return to the same system that they feed from (indirectly); people get access to model weights even if the entire model isn't open source. The same can't be said for OpenAI, Anthropic, Google etc (who also benefit from Chinese models and train on them).
Sure, the alternatives aren't a panacea of fairness but I'd much rather advocate for and support the thieves who give me a better deal if my choice is limited to thieves. Especially if thieves aren't hostile to their customers like Anthropic is (which is why I replied to you in the first place).
This was a problem for 5+ years ago. Nobody cares or at least the majority voice does not care across the world. Cat is out of the bag and there is no way to put it back in.
EDIT: Worth noting that I have long held the belief that if you put data out on the public sidewalk that you should have low to no expectation that it’s IP. It’s how I think about Google Maps data for example. If they want to reap the benefits by not walking it off the a user login than they can feel the pain if folks use that information. Same applies for media that has been bought, Reddit comments or any other datasets.
The difference is the Chinese models return to the same system that they feed from (indirectly); people get access to model weights even if the entire model isn't open source. The same can't be said for OpenAI, Anthropic, Google etc (who also benefit from Chinese models and train on them).
Further, Chinese models are significantly cheaper and the comapnies aren't hostile to their customers.
> Worth noting that I have long held the belief that if you put data out on the public sidewalk that you should have low to no expectation that it’s IP.
Except your beliefs aren't the cornerstone of modern jurisprudence. Why are models able to reliably produce replicas of Ghibli movies which go well beyond any example you listed?
(*) which are nothing short of amazing and are changing the world, there's no doubt about that.
> Thus in the eyes of OpenAI I do not exist in regard to SOTA security models.
I'm not seeing anywhere it says it's only limited to the U.S. Only that they had 'ongoing dialogue' with them. Which reads weird to me, how can an ongoing dialogue be past tense? But I digress.
> We’ve had ongoing dialogue with the U.S. government about our cyber approach, including today’s announcements and on our preparation for upcoming model releases.
> Third, even if I did, I will not be able to pass KYC because the typical KYC requirements are strictly tailored to a certain subset of the world's population and lifestyle choices, tuned by Americans according to their world view.
KYC is just that, Know Your Customer, if your 'permitted customers' are security researchers in the industry with a proven identity of employment etc then that is the KYC process, I don't see any issues with that.
> even if I pass KYC, my account will be banned by OpenAI immediately on the first prompt because they have close to 1B users and couldn't care less about any single one of them
Why do you assume this? Are you planning on intentionally trying to do something actively nefarious ? It's such a strange take.
Easy: it can be considered past tense in case "ongoing dialogue" is a corporatespeak for "f..k you". Which I believe is the case here. But that's an opinion.
> Know Your Customer [..] I don't see any issues with that
This might be the case if you're coming from a standpoint I have mentioned: the American one. This is a world view where everybody have physical paper documents proving residence, every labour effort is arranged in a very specific legal framework, every person have an address in a specific format, every person has one of just a few types of ID documents, etc, etc.
Problem is, the world have vast, vast differences in all of the mentioned areas and KYC companies couldn't care less because they are a business and they make money by KYCing as much people as possible for as little spend as possible. Thus they simply ignore any case that's not mainstream no matter how perfectly legal it is.
Being a digital nomad I cannot pass KYC at the vast majority of online services. My passport is under no sanctions, I do have residency in the first world country, etc., but passing KYC at Persona and others is not possible.
>> my account will be banned by OpenAI immediately > Why do you assume this?
Because of the risk profile. The company has no way of knowing whether "find all security vulnerabilities in this code" is a request from a whitehat or a blackhat hacker. The risk of someone using GPT to hack yet another DeFi project for a hundred millions while mentioning OpenAI is higher than perhaps a million user accounts, let alone a single one.
That's what KYC is for.
> This might be the case if you're coming from a standpoint I have mentioned: the American one.
I'm not in the US, nor America for that matter, I'm in the EU.
> Problem is, the world have vast, vast differences in all of the mentioned areas and KYC companies couldn't care less because they are a business and they make money by KYCing as much people as possible for as little spend as possible
"The lady doth protest too much, methinks".
There's not much constructive here other than a lot of assumptions and apparent malcontempt with how some businesses handle their business, but that's for another topic I think.
No, KYC has nothing to do with that problem. KYC doesn't help at all here.
> I'm not in the US, nor America for that matter, I'm in the EU.
Same here.
that's a bold statement. how does it not help solve the problem? what is a better solution?
How easy would it be for criminal orgs to setup legitimate looking fronts to pass these KYC checks?
How does it? Online KYC is a procedure to verify someone's documents and face. And that's it. What does it have to do with the actual usage of the OpenAI account and the code that is being examined with AI?
the system in place to prevent unauthorized abuse. by default, the guardrails are conservative. to reduce the guardrails you can jump through a progressive series of hoops to establish whether or not you have a valid use case. the entrypoint for establishing your use case is verifying your identity and background. if you don't want to do this, you are free to use Codex Security to identify and fix vulnerabilities, it is quite good at this. the harness and model are already evaluating the usage of the account and the nature of the code being examined and actions requested. but the again, the guardrail thresholds will be very conservative for anonymous users.
what is your proposal?
None. I don't see a solution.
I'm silently rooting for Chinese models here.
I guess you would also not provide KYC to a bank that provides number additions/subtractions between database rows services
And I'm happy to pass KYC in person, so I do have accounts in different parts of the world. It's the online KYC that's not passable for digital nomads like me.
Its built-in resume mechanism didn't work after it crashed when running out of my 5 hour session limit, but Claude Code was easily able to resume it 5 hours later reading the session logs and https://openai.com/codex/security/scan.sh
It certainly has nothing to do with openAI's co-founders donating to the current administrations election fund, are actively supporting the DoW war efforts of autonomous weapons and also otherwise being ideology tightly coupled with the current US government.
We don't know that it is Mythos level, it could very well be at (guardrailed) Fable or below.
This is not a wide open distribution, this is only being provided to hand picked partners, similar to how Mythos was distributed (unlike Fable which had wider distribution)
The larger question, which I don't see an answer to in this post:
1) was this tested and validated by the US Government?
2) is the list of partners vetted by the US Government?
If This is "mythos-class" AND
OpenAI approves SK Telecom as a trusted partner ( https://www.wired.com/story/sk-telecom-anthropic-mythos-export-controls/ )
OpenAI did not get approval.
Main point being, we have no idea the measurable capabilities of this, it could be as great or better as unrestricted Mythos, or on par with guardrailed Fable.. or just OpenAI hype that measures up to neither.
The distinction is important because if it truly is a Mythos level (or even guardrailed Fable) it in theory would require that 30 day US government validation before release as well as oversight to the approved partners allowed to use it.
Op was drawing a parallel as to why we should be outraged at the double standard, I was drawing a better parallel by which to compare.
Anthropic has successfully made their vibe the bringers of the end of the world so please stop us. ChatGPT is in some weird middle ground where they’re no longer the evil company of a year ago and are trying to act like the mature one in the business. XAI is the vanity project of a jealous kid who wants to ignore anyone elses opinion but is falling behind. Gemini is viewed as the less cool cousin and no one pays attention to since google has generally avoided the end of the world talk (coincidentally I think gemini has the best non-coding model but no one seems to talk about that)
I think each company has been treated in the wider discourse based on their vibe and no one is operating on facts here since it’s all pretty obfuscated intentionally. I could see some things going very bad as a result of these models but I also don’t believe anyone who’s just pattern matching to their favorite scifi book here and we’re all playing into it. May as well go outside for a bit and enjoy the sun.
Most things should be fact based, but you also should have limits to your marketing
Should a nuclear energy startup that says "nuclear weapons are very dangerous and should be regulated" be liable to have its assets frozen?
DoD signed a contract with Anthropic that said it can't use Claude 1) to run a fully-autonomous killchain or 2) to surveil Americans.
DoD then decided they did not want to abide by the contract they signed, and they've since launched a pressure campaign to coerce Anthropic into reneging on the terms they already agreed to.
That doesn't mean it matches Mythos ability to write functional exploits.
Maybe it is, maybe it's not. They don't tell.
The problem is, though, given Anthropic have said all of that, they really have very little grounds for objecting to the US government's intervention here. Everything that the government would have to prove to justify their intervention has already been freely admitted by Anthropic, even though the "admission" was maybe more intended as a marketing ploy.
OpenAI, four months ago, started to require users to verify their identity if they flagged their activities on frontier models (gpt-5.3-codex and higher) as risky. Their filters were originally quite coarse and it resulted in a ton of normal tasks being flagged. There was a lot of drama about it at the time, but it seems like things have smoothed out.
KYC goes back to a year or two ago. API access to gpt-image-1 required it.
There is no doubt in my mind that he is personally pissed because Anthropic stood up to him.
That makes sense if both openai and anthropic have export restrictions on their similar models. If they didn't then it seems like the comment you're replying to may be correct.
Mythos is a serious model; the NSA said it compromised them in hours.
This is a marketing article.
This quote from Pratchett pretty well summarizes Anthropic marketing approach. OAI are quite more restrained in advertising their model's doomsday abilities which greatly diminishes the pretexts that USG can use.
Two things can both be true - that Trump's administration is petty and vindicative and that Anthropic are reckless and not doing their best in winning hearts and minds.
They're releasing ever more powerful models with stronger offensive capabilities. So do they have to help bolster the defense of all existing software, just... forever?
If we advance both the offence and defense with each new release, is this sustainable?