I see a lot of knee-jerk comments to this, but I highly recommend running a scan ( https://openai.com/daybreak/codex-security-plugin/#codex-cli ) in your projects so you can evaluate it yourself. It found a real security issue in a project of mine, with very few false-positives.
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
I don't know what the solution to this is, but I find it somewhat unfair that I pay money to Anthropic, and I pay money to OpenAI, and neither of them will let me use their best models for securing the software I work on.
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!
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.
The problem is even worse than that. OpenAI and Anthropic have your source code and superior knowledge of its vulnerabilities. All you can do is hope that they won't one day use it against you.
I think using open weight models will solve this. I believe they are nearly caught up and much of the gains are in the harnesses or properly orchestration of subqueries. (I'm no expert, just my opinion).
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.
A large part of vulnerability analysis is just having the time to crunch through enough possibilities. Expertise and smarts definitely speed this up but there's a lot of just turning the crank until something falls out. Even a relatively dumb model with some good prompting will fine vulnerabilities if you ask it to and give it the time and resources to do so.
Pretty sure the secret sauce is in the summarised thinking.
Maybe better though process… But I have a feeling it’s server side tools and a scratch space to prepare the reply.
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”.
> 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 sure I follow your logic. Paying for a service does not mean you get access to all potential services a provider offers. Providers can choose to keep some services internal.
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.
No one commenting on the fact that oAI is releasing a Claude Mythos-class model - with apparent 0 restrictions or concerns by the US government, while Anthropic's (their competitor) model has been pulled weeks prior by the administration for 'security' reasons.
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.
>No one commenting on the fact that oAI is releasing a Claude Mythos-class model - with apparent 0 restrictions or concerns by the US government
We don't know that it is Mythos level, it could very well be at 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/ )
OR
OpenAI did not get approval.
will this be shut down as quick? Otherwise, it is not really a comparable scenario.
Man, some of you will invent conspiracy theories to justify some deeply cynical fiction. OAI has been more proactive about doing customer KYC than A\.
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.
Oh! So the new openai model is limited to US residents and they use their existing KYC process to verify it?
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.
Do you think that Anthropic's models would have been pulled if they did not say for months how their models is basically going to break the whole internet and that governments should most definitely restrict AI? I doubt it.
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.
Ok so why I don’t have access to this if I already pay for the max plan? Should I pay a security researcher to run codex on my code? Is this how it is supposed to work? Let’s hope we get some real cyber models that people can actually use from the Chinese without the stupid application forms.
Why do you think you should have access…? People who pay enterprise API rates also don’t if this makes you feel better (it shouldn’t, you shouldn’t have felt bad in the first place)
This is how you do it when you're not AS childish. You go "here's a model for cybersecurity" and put a price on it. I know they're releasing it to some vendors first, etc. but the lack of a clown spectacle is nice.
The whole "it's too dangerous to release!" is complete hogwash.
A person can take a hammer, walk out in the street, and we can count how many people he can kill with the hammer before he is stopped. My local hardware store still sells hammers, and I haven't seen the CEO of it claim that their hammers are much more dangerous and it's totally going to end the world if you allow any random person to have one!
If that hammer could allow people to go into people's homes / work en masse, steal all their information, blackmail them, steal their identities, break their systems (including those of hospitals and other critical infrastructure) and generally help fund bad actors through it all we'd think of having restrictions on hammers too. A hammer can't screw people over by the millions.
I don't like this argument specifically with AI. Facial recognition everywhere you go is just a tool. Your job creating a detailed profile on exactly how you work, who you talk to, and about what is just a tool. The tools have become so good and easy to use we have to have serious discussions about them before things get out of hand.
Did you see how close the non-sheltered available models come? They come quite close. Most people aren't even using them for this purpose, but they could, and this is our reality. This is why your argument fails.
Disagree. @lionkor compared them to a hammer, and @ragequittah is saying they're not like a hammer.
The narrow gap between downloadable and frontier models is tangential to this. If you want to expand on the "hammer" metaphor, the downloadable models are a small construction/demolitions firm, and the frontier models are a big construction/demolitions firm.
In this analogy, there's no training school or certifications for the staff either of them hire, and society is still working out what public liability requirements and planning permission laws are even though both companies are being hired all over the place, because everything they do was only invented a few years ago.
So the solution is... giving up? Let the technogods do whatever they please? Because we are not talking about storms and earthquakes, but about humans in power.
The risk of catching federal charges, proper jail time and aggressive responses from law enforcement is a far more effective means of preventing malicious behavior than anything proposed so far.
I can go into stores that sell things that are much more dangerous than hammers (or frontier cyber models) and no one will give me a hard time about it.
It's toxic to call out big companies fearmongering about how their AI is too smart to be accessable? And it's somehow comparable to telling newbies asking question to RTFM?
It's a pretty interesting opportunity. I wonder if they will reach to companies and tell them how many things they could fix and how many are critical, before selling them the solution.
Interesting release. I’m most curious about how well this holds up in messy real-world environments, since that’s usually where specialized benchmark gains get tested.
Does the EU CRA now mean that every European company that either sells software or sells anything that has a software component is now forced to pay for this by September and update their software?
I think if nothing happens from the government, then this would be a very good example of the benefit of keeping your mouse shut especially if you are lying to get some hype like Anthropic did for months.
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
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!
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.
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.
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?
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”.
> 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
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.
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 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
OR will this be shut down as quick? Otherwise, it is not really a comparable scenario.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.
https://openai.com/index/trusted-access-for-cyber/
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.
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.
The whole "it's too dangerous to release!" is complete hogwash.
A person can take a hammer, walk out in the street, and we can count how many people he can kill with the hammer before he is stopped. My local hardware store still sells hammers, and I haven't seen the CEO of it claim that their hammers are much more dangerous and it's totally going to end the world if you allow any random person to have one!
I don't like this argument specifically with AI. Facial recognition everywhere you go is just a tool. Your job creating a detailed profile on exactly how you work, who you talk to, and about what is just a tool. The tools have become so good and easy to use we have to have serious discussions about them before things get out of hand.
The narrow gap between downloadable and frontier models is tangential to this. If you want to expand on the "hammer" metaphor, the downloadable models are a small construction/demolitions firm, and the frontier models are a big construction/demolitions firm.
In this analogy, there's no training school or certifications for the staff either of them hire, and society is still working out what public liability requirements and planning permission laws are even though both companies are being hired all over the place, because everything they do was only invented a few years ago.
Like, e.g. the USACE
I can go into stores that sell things that are much more dangerous than hammers (or frontier cyber models) and no one will give me a hard time about it.
1. Browse the internet
2. See what people hate about OpenAI
3. Adopt the worse version of it
4. Profit?
Sam Altman fearmongered about AI alignment - we fearmonger harder.
OpenAI is CloseAI now - we are even less open.
OpenAI is going to IPO - we IPO first.
Really?