Due the the size of the default system prompt you get over 100k really in easy in Claude, and because that is fixed 200k is really almost 3x more room to work with, not 2x.
200k also used to be the limit for Sonnet for a long time.
If Haiku 5.5 also had 200k, it'd actually be usable for us. As it is, it's Luna or Sol.
On purpose? I was using qwen3.8-27b the other day and very quickly ended up burning through half of the context (it has 128k total) pretty quickly. Not because of my directions, but simply because the codebase was very large.
256k seems sweet spot. Even large codebase can be handled you just has to map out areas for agent. Anything past 256k usually exception than norm to be.
At least this my discovery when I did evals on my reviewers skills. 10k LOC is high end for high effort or around 150k context using opus. You can push towards 15k but that’s on max effort with about 400-500k ctx.
Fast models like haiku do not even need 1M ctx in my view
Yeah codebase structure is very important. Admittedly, by large I mean average size but very poorly structured with a spaghetti (and sometimes cyclical) import graph.
1. First I hear about it is a heads-up at 5:00 am 21st.
2. Available info then is pretty much, "omg, it's the end of the internet".
3. There are no specifics available on exploitable configurations.
4. Updates are not expected to be available for at least two more hours, so people start deciding to do the Madagascar "shut down everything" meme.
5. Cloudflare says they're seeing exploits in the wild, well, that's not great.
6. I shut down everything, including a national brand, because wtf else am I supposed to do at this point?
7. Updates become available at 0645. Fortunately, I don't run my entire infra on Docker, so I'm immediately deploying the updates to all configurations and bringing sites back online approximately in order of traffic volume.
8. Still no specifics on the exploit.
9. A million vibe-coded PoCs appear on GitHub ... and none of them work.
10. It's late morning local time on the 21st and I'm still bringing things back online, having taken the day off from my regular job.
11. Conversation starts to shift towards, "...oh, did we over-react? How exploitable is this exactly?"
12. Still no specifics available.
13. Somebody on the internet gets as fed up as I am with the whole chaos vortex of idiocy and stands up a site saying, "okay, here's some money if you can exploit it."
14. Hours pass.
15. Still no specifics available.
16. Zeitgeist becomes, "I guess it's not exploitable after all."
17. The test site gets 'sploited.
18. Finally somebody actually posts the damn specifics -- namely, the person that grabbed the $2k.
19. Late morning local time on the 26th, and the conversation becomes, "well, obviously this was exploitable!" <-- we are here
Coincidentally, I just started reading Crisis Engineering, which seems like an excellent and applicable book, but I'm not sure who to beat with it.
It’s worth noting that while folks were getting busy with 11 and 13, plenty of experts in the field were saying ~“yes, we do have viable exploits here, but we’re not sharing them at this time. This is a big deal.” And that the person who spun up 13 spun up a static html page at first and also misrepresented how vanilla the stack was.
We could have skipped from 7 to 19 if folks had just taken seriously the guidance from experts in the field.
I mean, it's clear you have a valid reason to be irritated, but that's not the concern I have here. What's worrying is people applying motivating reasoning to talk themselves (and others) into believing the vulnerability simply isn't a big deal, which is clearly false.
Yeah. I'm choosing to believe that those people just over-weighted some previous info and are under-weighting the latest, but I dunno.
There's not a chance in hell that even half of the world's PHP has been updated, so I would expect that if we start seeing large-scale exploitation, those people will change their opinion pretty fast.
Being totally burned out on AI hype myself, I'm somewhat sympathetic towards the folks that are still thinking of this as just another example of that.
> Being totally burned out on AI hype myself, I'm somewhat sympathetic towards the folks that are still thinking of this as just another example of that.
I mean, it was also that it was marketed in the media as a security apocalypse and unpatched servers would be falling within 15 minutes after the patch got released. This didn't happen. And that definitely warranted skepticism. There's been MUCH more severe vulnerabilities in the past that just went by without OpenAI's marketing making everyone go crazy about it.
Also, almost all information about it was dressed up in slop. Because OpenAI didn't want us to forget that AI was involved, somehow.
> believing the vulnerability simply isn't a big deal, which is clearly false.
you keep repeating this, also when people make very valid other criticisms, such as the hype causing government websites to be pre-emptively shut down -- do you believe this was warranted?
because that is the criticism.
nobody, including that "PHP vuln isn't so bad" blog article, which was VERY clear about this, said you shouldn't patch your servers
what's worrying is that you keep making up straw men, but why?
this isn't about the vulnerability, it's about OpenAI's Patch The Planet hype train
No, but they were part of the PR blitz. The lack of real-world exploitation + Calif commissioning a fucking webcomic made me skeptical. That's not fucking normal. Given what a huge marketing stunt it turned into, and 1000 vibe coded PoCs that aren't PoCs, it was at least in the realm of possibility that Calif huffed their own vibes too hard. If you want people to take security risks seriously, making them into a media spectacle promoting OpenAI is not the way to do it.
It, along with the press releases and the cutesy domain names registered, shows that this was clearly being played up as a marketing stunt, with planned coordination of such ahead of time. As soon as you begin to use the security incident to communicate a coordinated narrative about how amazing AI is, it's the most given thing in the world that some people will become skeptical of the severity of it and whether it isn't being played up as the most impactful security vulnerability ever found (per the press release claims) solely as part of an attempt to make the AI company sound more impressive.
AI company tells you their AI was used to find the biggest vulnerability EVER, with much fanfare. Zero cases of exploitation in the wild. Zero publicly demonstrated PoCs. It would almost be weirder to not be skeptical at that point unless you for some reason trusted everything a marketing department says implicitly.
> These findings rank among the most severe vulnerabilities discovered in PHP, and arguably in software history.
See also the parable "Boy Who Cried Wolf" for more insight into the human condition.
I suspect so, and just a general blowback effect where if what people perceived to be the message is even slightly more dramatic than the reality then there's tendency for many to overshoot reality in the other direction.
I mean what is the world coming to when security researchers cannot release a writeup that is designed to look like slop, and starts with a bunch of cartoons with lame cyberpunk and matrix references, without being ridiculed for taking the piss ...
Not long ago the opposite would be true and hackers would be ridiculed by their community for being too sterile and corporate, and everyone else would be ridiculed for over-indexing on playful presentation as opposed to the meat of the technical issue.
It’s been going on for far longer, people want to believe all OpenAI and Anthropic are doing is pure marketing. When the first Fable results were released, everybody was skeptical (which is rational), but with every third party confirmation of the capabilities they doubled down that this must be even more marketing instead of admitting that they were wrong. To the point that they are no longer skeptical but in pure denial about what is happening.
A bunch of folks tried to say the challenge was dubious at best and that the vulnerability was a legitimate concern, and were told by commenters here that they were fearmongering or shills.
Evidence has shown they were pretty spot on about the first part (the test originally served static html, the php code wasn’t actually as vanilla as described, and the bitcoin key wasn’t actually there).
I’d have hoped this kind of clear vindication on the first point might have caused folks to revisit their stance on whether to trust the expert stance on the second. But maybe I’ve been overly optimistic.
Where do you get the two hours from? Did this attacker spend two hours on it, after 17 hours of others failing to break in?
> The fact that people are still in denial over this is crazy.
Maybe you misunderstand what I say.
Imagine the tens of millions of personal or small business websites running php. What's your guess, what portion of the unpatched ones is already pwned? Is it 1%? 5%? 15%? 50%? 95%?
I have working RCE as long as the server has a memory_limit that is less than the max POST size. From my reading of the PHP patch, there's no way around that part, but a lot of real sites are running like this.
No idea of course if this meets the authors' "out of the box" condition. Who runs a fully stock PHP config?
I don't think it's implied the wrapping keys were the weak ones. More like, the script that ran the wrapping still had the fast "debug" generation of a pair of keys to test the wrapper on enabled, for example.
If your generation is any good the odds of ending up on such a number should be astronomically small.
Biasing something that should be random tends to badly backfire. See the Enigma who avoided enciphering letters to themselves, which was then used to break it.
There's very few stuff that needs Astra/Opus level intelligence.
Luna 6.1 would be amazing.
reply