443 points | 2d ago | Discuss on Hacker News | Back to Radar
A lot of similar aspects to the "PHP embargo was broken" deal the other week.
For the same reasons you give, there is a different level of pressure on timelines and responsiveness, too, for a project like OpenJDK. I don't think 1 month is reasonable timeline for a publicly known vulnerability like this.
Not all vulnerabilities are created equal, and they don't all justify an emergency release (the JDK has regular quarterly releases for security patches). Until you know the impact, disrupting your users is not the best path to security (just as you don't want to police to respond to every suspicious occurrence).
> I don't think 1 month is reasonable timeline for a publicly known vulnerability like this.
It wasn't a publicly known vulnerability; it was a publicly known bug. If the original reporter had thought this to be a vulnerability he wouldn't have reported it like that (he's a former colleague who's familiar with project policies).
You might follow up with a De Bruijn/cyclic pattern and can see the overflow happen by step1 and precisely locate it by what replaces the A's.
https://youtu.be/L2j90atBpRU?is=q6W_1eF4viMoKZEA
Didn't the Alpine Linux bug also use A's for the exploit to login as root?
- why was this not a secure vulnerability report, why was it public?
- why was this closed [by the submitter]?
For the second question, I suspect it's because they were told a different fix was coming.
I know I have opened a lot of PRs where the only result is I get bot spam in my email and no project reaction, so I close the PR to stop the bot spam; if nobody from Oracle has reacted in 2 months it seems unlikely anyone will.
Nowadays I’ve stopped opening “cold-call” PRs like this. I find a way to chat with a maintainer first and anchor what I want to do (“I hit this bug, would you take a PR to fix it? I was thinking it could be solved by x, would that work?”
Much much better outcomes that way.
Because it was seen as a crash not an RCE. AI's power has been chaining simple crashes into exploits. But that's not generally going to be true.
There are lots of bugs in lots of software, and lots of them get ignored when reported.
I personally aware of multiple security vulnerabilities in major pieces of software that everyone has heard of, which I very quickly and easily stumbled on while working on this software as an AWS engineer several years ago. I had never used an LLM of any kind at the time.
I presented my findings both internally and externally, and the reactions ranged from shrugs to delays to hostility towards me for "getting distracted" or "being overly negative."
Those bugs that I discovered still exist, last time I checked.
the contrary is very likely thing the agent may have done, given they... like agents dig the web, don't they. and if im a security researcher digging through Oracle's shit, the open issues sure does look a good starting point.
Maybe simply a coincidence, but I would put my bets that the first person with the PR was also using Codex, and the model was exposed to the details of the bug.
I have no idea why people keep talking like this. Probably because sadly most engineers lack the ego to back themselves up. Imagine if someone said this in 2015: "I worked on the physics engine, tests, via widely accessible libraries. Well, more precisely, I told my compiler to read the source code and compile it like any serious software engineer in 2026."
Quit saying Codex/Claude/an LLM did something. They didn't. You did.
I think there's a vast difference between giving source code to a compiler (even a superoptimizing compiler) and giving a prompt to a language model. This distinction absolutely deserves to be called out and not swept under the rug. It doesn't make you any less responsible for what you submit, but it's courteous to your reviewers to be transparent.
There is a massive difference between prompting "security audit this codebase, make no mistakes", performing an in-depth agent or LLM-assisted focused search for security relevant code, or otherwise manually reviewing code as we did in the pre-LLM era. It behooves us to be honest about our own personal contribution to our output - both towards others and to ourselves.
Your hypothetical makes the opposite point from what you intend. If someone in 2015 said that they worked on a physics engine, and all they did was run the compiler on existing code, we'd laugh them out of the room. Pointing an existing computer program at existing work doesn't count as doing very much.
Because when people really spend the effort required to fully review and own the result, somehow they don't have a problem of pointing at models.
It was literal bug with provided fix. All it needed is someone to have a proper look.
But someone clowned a bot under a visionary thought "if nobody looked it must not be important" and here we are
> The stub stores past the logical output length before returning to Java, corrupting the heap.
It was obviously a security vulnerability. That the project didn't immediately re-write then close the PR and admonish submitter is astounding.
The takeaway from this is not "the submitter was wrong", it's "we need to move off of openjdk yesterday".
Did OpenJDK fix the bug only when Java4shell came out, or did they have a fix in progress? Gemini and Claude tell me the fix had already been in the works. Therefore your earlier:
> > > Given they (those best able to understand the significance of the crash) didn't consider the bug worth fixing,
is false, and indeed that was 'facts not in evidence'.
> If the submitter went through project zero we'd be well past the 90 day window. How is that not sufficient evidence?
There is no evidence of "they [...] didn't consider the bug worth fixing". On the contrary, there is evidence that they did consider it worth fixing.
Sorry, but why?
Why should someone discovering a bug in an enormously complex piece of software have any responsibility for assessing the security implications of that bug?
If the maintainers wanted to make it private they could have done so.
> The stub stores past the logical output length before returning to Java, corrupting the heap.
That's obviously a security vulnerability. If they didn't understand that, how did they even find it?
Why invest more time to close it? The automation will do it for you.
In this case, there's evidence from a linked issue that Oracle was in communication with the author outside the PR. When you get a ping, and you know it's being worked on outside the PR, it does make sense to close it, cause you know it's being resolved elsewhere.
If I were the relevant maintainer I would have reached out to submitter privately and asked them to hold off more widely reporting this, and I would not have stopped the bot from announcing impending closure.
Would it be better to have some Oracle gpus running the dukebot? Are the robobun and friends better at real triage/escalation? Could we automate the robobots to register the `java2root.com` domain on its own, release a blog post and post it in twitter/hn upon every successful security issue?
Except soon you can expect LLM triage to be standard.
We would then find out if enough tokens would be invested in this one to figure out it was important, but any look is better than none.
I actually think the suggestion in the PR should have been auto accepted by an LLM.
We're not talking any large open source project here. We're talking about a project that, like Linux, runs much of the world's software infrastructure, and even bugs that could potentially be vulnerabilities need to be handled with great care.
Also, it's not really a leaked 0-day because there's ways to go from a bug to an exploit (even going from Calif's detailed public report to an exploit is not trivial, because exploits depend on the particular application; there has been no disclosed exploit that requires only the JDK).
I'm guessing Oracle won't get back to the person later and say "thanks now fixed was actually a high sev vuln" (I can understand if everyone is short of time and can't keep track of such things?)
Nonetheless, being able to understand the criticality of a finding, and how to exploit it are other skills, that still require supervision.
I often do this when looking for low hanging fruit. A lot of the time you find half-abandoned PRs when the original author ran out of damns to give.
This PR's author says there on October 1st: "Got an email from Oracle asking me to NOT proceed with a PR until receiving approval." and then on October 28th: "It's taken over by Oracle. Nothing required from us."
https://github.com/adoptium/adoptium-support/issues/1480#iss...
edit: also oracle is big enough to have one of every modern gen to test on
They are the biggest asshole of a company when it comes to anything, but especially security issues. They are one of the least friendly companies to security researchers and are likely to attack you with lawyers.
Fuck Oracle. We should be putting them out of business.
The sign: "Do not fall into the trap of anthropomorphizing Larry Ellison. Think of Larry Ellison the way you think of a lawn mower. You don't anthropomorphize your lawn mower. Your lawn mower just mows the lawn. You stick your hand in there, it'll chop it off, the end. You don't think "oh the lawn hates me!" the lawn mower doesn't give a shit about you. Your lawn mower can't hate you. Don't anthropomorphize the lawn mower. Don't fall into that trap about Oracle."
Because the company didn't start as a CIA domestic surveillance project? CIA project, yes. CIA "domestic surveillance" project, no. The CIA, like any large org at the time, needed a way to store large amounts of data in a usable form.
It's weird how much this type of conspiratorial thinking has become commonplace here and in other tech discussions online. I'm not sure if there's just echo chambers people fall into where this is repeated enough that people assume it's true?
If every bug was a zero day then we'd have never hooked things like water grids and electrical grids to networks. Grocery stores and bloodbanks would have never put freezers on the internet. Along with a million other huge security risks, too many to be named here.
The AI age has exposed that stupidity.
Capitalism rewards all sorts of safety requirements, if they become visible to a balance sheet within a reasonable period of time. Just look at marine insurance in the middle east right now.
The status quo is that IoT freezers are massively inconvenient every decade or so. Businesses can live out an entire lifecycle in this period of time, without the risk ever hitting their balance sheet.
pixl97 is saying: what if we made the risk far more common, so that its effect shows up on a balance sheet within 6-12 months?
What is the purpose of this? It's just intentionally making hacks more common and a net negative for the economy.
More secure systems aren't inherently good. You don't hear people talking about how happy they are that break-ins are common and everyone locks their doors before leaving the house.
I'm sorry buddy, but you were born in the wrong universe not to defend yourself.
The internet is a global network. This means people not under your sovereign jurisdiction can reach out and fuck with your data and you can do jack shit about it unless you want to take a bunch of other really expensive actions (for example economic or hot warfare).
Computer people should know better than thinking that the internet is anything other than an inherently hostile place. Groups we consider enemy nations share the same web space. Countries that make $1000 a year share the same web space, and all the juicy servers you leave insecure look like a very good way for them to become absolutely rich. Hell, even people in your own country are apt to get away with hacking with a little bit of skill.
The internet is not your neighborhood. You can't call the police to fix every little issue for you. It is a literal war zone, just most people are still in denial about that.
Now start treating it like that.
Please feel free to explain how people are not aware of the danger of running over their own children because of the daddy state, and how fun owners have somehow missed the dangers of guns.
The consequences are usually suffered by other people. You speed, other people die in the resulting crash. You get the freedom to make unsafe buildings and furniture and electrical goods for profit, other people die from them collapsing and burning down.
DATASET
Underlying Cause of Death, 2018-2024
CONDITIONS:
age >= 1 AND age <= 14 GROUP BY 'ICD-10 113 Cause List'
+--------------------------------------------------------------------+------+---------------+
|ICD-10 113 Cause List |Deaths|95% CI Crude |
+--------------------------------------------------------------------+------+---------------+
|#Accidents (unintentional injuries) (V01-X59,Y85-Y86) |19,632|5.0 (4.9 - 5.0)|
| Nontransport accidents (W00-X59,Y86) |10,596|2.7 (2.6 - 2.7)|
| Transport accidents (V01-V99,Y85) | 9,036|2.3 (2.2 - 2.3)|
| Motor vehicle accidents (V02-V89...) | 8,550|2.2 (2.1 - 2.2)|
|All other diseases (Residual) | 8,151|2.1 (2.0 - 2.1)|
|#Malignant neoplasms (C00-C97) | 7,750|2.0 (1.9 - 2.0)|
|#Congenital malformations, deformations & chromosomal abn. (Q00-Q99)| 5,564|1.4 (1.4 - 1.4)|
|#Assault (homicide) (*U01-*U02,X85-Y09,Y87.1) | 5,225|1.3 (1.3 - 1.4)|
|Accidental drowning and submersion (W65-W74) | 4,700|1.2 (1.2 - 1.2)|
|#Intentional self-harm (suicide) (*U03,X60-X84,Y87.0) | 3,839|1.0 (0.9 - 1.0)|
|Major cardiovascular diseases (I00-I78) | 3,271|0.8 (0.8 - 0.9)|
|Other & unspec. nontransport accidents & sequelae (W20-Y86...) | 2,717|0.7 (0.7 - 0.7)|
|Symptoms, signs & abnormal clin./lab findings, NEC (R00-R99) | 2,658|0.7 (0.6 - 0.7)|
|Malignant neoplasms of meninges, brain & other CNS (C70-C72) | 2,650|0.7 (0.6 - 0.7)|
|Assault (homicide) by other/unspec. means & sequelae (*U01-Y87.1...)| 2,625|0.7 (0.6 - 0.7)|
|Assault (homicide) by discharge of firearms (*U01.4,X93-X95) | 2,600|0.7 (0.6 - 0.7)|
|Intentional self-harm (suicide) by other/unspec. means (*U03-Y87.0) | 2,452|0.6 (0.6 - 0.6)|
|All other & unspec. malignant neoplasms (C17-C97...) | 2,435|0.6 (0.6 - 0.6)|
|Other & unspec. infectious/parasitic diseases & sequelae (A00-U07.1)| 2,177|0.6 (0.5 - 0.6)|
+--------------------------------------------------------------------+------+---------------+Well 'frontovers' (parent drives massive SUV over child because of vehicle's poor visibility below the hood) only kill 366 of the 4074 children killed in motor vehicle accidents every year, not most of them.
From 2026 - Letter to the editor, New England Journal of Medicine: """Previous analyses of mortality data from the Centers for Disease Control and Prevention (CDC) showed that the leading cause of death among children and adolescents (defined as persons 1 to 19 years of age) shifted from motor vehicle collisions to firearm-related injuries in 2020.""" - https://www.nejm.org/doi/full/10.1056/NEJMc2600445
Why is there not a proxy so we don't need to expose the refrigerators to even a VPN?
There's no real problem being solved. People just went that direction to follow trend of 'smart' everything because it's cheap, you can concoct hypothetical scenarios where it might be useful for marketing (regardless of whether or not they're at all realistic), and most importantly of all - complete lack of awareness about the risks involved.
There actually is. Humans are notoriously sloppy in following mundane, boring compliance bullshit tasks, and noting down freezer temperatures is what I'd say very high up on the list of bullshit tasks that are so pointless they're demeaning to the human who has to be the eyes for the machine.
Connecting the freezers to a central logging facility now needs no human checking the freezer every hour and writing down the temperature on a sheet of paper no one will ever read. It is precisely what machines are made for. And since compliance is now automatic instead of a strong incentive for the human to just invent "valid" data and be done, actual outages can now be spotted much faster.
Getting informed about outages or pending outages (e.g. by not just monitoring target temperature but also motor temperature, current consumption and vibration) is just a nice cherry on top.
I don't think that's true.
They may have known there was a risk, but historically it was small. If they knew it was certain to be hacked before next Thursday most would stop doing it pretty quickly.
It's like driving. Every time you drive there's a small chance of an accident, but everyone still does it. If there was a 50% chance of death most would stop doing it pretty quickly.
Similarly, (securely) networking water treatment plants enables people to remotely verify that settings are where they're supposed to be and adjust them to match changing sample reports or intake conditions, reducing the risk of over-treating or under-treating the water, both of which can have safety consequences.
On top of that, networking enables much greater efficiency. Rather than having a person sitting at every substation ready to throw a breaker, or more likely, ready to drive there (taking time), you can have them sitting in a control room that monitors the whole grid at once.
Much of modern life would be much more expensive and much less safe without networking.
We really shouldn't throw the baby out with the bathwater. It is possible to build secure systems. It takes a certain degree of paranoia and dedication that humans find difficult, and a willingness and ability to respond quickly to unknown threats, but that doesn't make it impossible.
The modern world wouldn't work without gasoline, but it's burning the planet to the ground in the meantime.
If your efficiency is gained by taking the brakes off the car to make it lighter you're going to realize how bad you screwed up when it comes time to stop.
We've been playing a 30 year game of fucking around and we're reached find out time.
If we want to build secure systems we're going to have to flush about 95% of the programmers out there.
I agree about one thing - incentives cannot work properly if it takes you 30 years to find out but your economy is built on 3 month business cycles.
We don't need to flush 95% of the programmers (though that is probably happening anyway). We need to flush 95% of the incentives to go fast and avoid spending money. Fast, cheap, or good - if you're very lucky and very good, you can squeeze two of them out of the equation. Most of the time, you're lucky to get even one.
I don't know if people here understand how these things need to be done in the most important and critical open source projects, such as Linux or OpenJDK, but there are two things you must not do when an issue is suspected to be a risk until you've researched the matter and decided on a backport and release strategy: You shouldn't communicate publicly, and you absolutely shouldn't merge a patch.
Full disclosure might cause short term panic, but it does get things fixed quickly.
I (still) agree, but I don't think these rules will be valid for much longer, though.
With new models capable of finding vulnerabilities faster than ever before, if you (or your AI) was able to find a critical bug, it's only a matter of time that the bad guys will also discover it - if they haven't already.
Human skill and ingenuity used to be the limiting factor, so it was the universal equalizer. Now the game depends on how many tokens you want to throw at the problem, and there are many bad actors with incentives to outspend every white hacker out there.
Responsible disclosure will still be important, of course, but timelines will need to be compressed into days, instead of months.
It's expected that the ability to run models follows a power law distribution: very few actors are able to run the most expensive models on many codebases/binaries, spending lots of tokens on each project.
Disclosing a serious vulnerability potentially exposes it to orders of magnitude more people.
If I spent N tokens to find a specific funky bug, you should expect others to have done the same. In a way, in this modern age the full-disclosure guys had it right all along. Reality caught up with a three decade lag.
Shit like this is what forces people to just _merge_ fixes, why shouldn't they?
So until you know how dangerous the bug is and have a plan for release (does it go in the next quarterly patch, which users are ready for, or are you making a special release?) you shouldn't do anything public because you don't yet know what to advise your users to do.
Exploit shops can start months later to evaluating exploitability.
Timeline was:
- Jun 8: reported crashes
- Sep 9: (two months!) it was know base64 asm code is bad and patch was posted
No idea why anyone need to wait for anyone and especially on any corporation to apply the available patch !
So, it's 2 months later and Oracle is siting on their hands.
Then month and two weeks later Calif rediscovered crashes but this time person(s) involved have funny past time hobby: hacking :)
5 days later Oracle looks clueless, requesting info...
Two more weeks later Oracle have The Patch ! And recommends everyone to do nothing !
14-17 Nov: governments are informed :))) Thank you Calif for remembered about them ;) Somebody needs to tel governments they have problems...
17 Nov: Oracle gives fixes to the world...
Now governments are allowed and can start patching their infra...
Is any of this sane ???
It's almost half a year of exploitable Java on government and enterprises servers and fix was ready by half of that time.
Just nobody was allowed to compile open sources by themselves.
And all of this is optimistic version, thanks to code being available...
> No idea why anyone need to wait for anyone and especially on any corporation to apply the available patch !
Because as soon as it was suspected to be a security risk we could not apply it. This is how security is done in critical products.
When you say "accept the PR", understand that we're not talking about applying it to mainline, but backporting it to all affected and supported versions. That is something that immediately draws attention. Then, a vulnerability is closed not when it reaches the codebase, but when it's deployed by the users to production, and users have the means to apply the regular quarterly updates, but not necessarily emergency ones (because there can always be other emergencies that occupy them). So what you need to do is investigate the risk and then decide whether to issue an emergency update or wait for the regular quarterly release.
The JDK is critical infrastructure, we simply cannot take cavalier, irresponsible actions of the kind you suggest; I would say insane, but I realise you're just unfamiliar with critical software security (I know you have good intentions and think it's fast, but we've been doing this for thirty years and know it's anything but). Think what would happen if every critical product would issue an emergency release with the maximal urgency advisory whenever any bug is thought to might be a vulnerability of an unknown severity.
Sorry, but we must take security matters extremely seriously, and handle them the right way (which is also the fastest way).
Although maybe they didn’t even know their LLM was doing this.
The harness itself doesn't run with Internet access and I confirmed it didn't access this PR directly, but who knows if this was in the training set itself.
Either way, these simultaneous disclosures are incredibly common now. The Ruby issue affecting Gitlab was not only known about for 6 months, it was publicly fixed for 6 months before someone turned it into an RCE. The fix sat on the Ruby dev branch, and it was never pulled into an actual Ruby release until after Gitlab disclosed their own 10.0 vulnerability through it.
It's very common for reporters to not realize the impact of a bug, models are very good at turning bugs into RCE chains, and the volume everywhere now means security teams are basically only focused on RCEs.
I’m curious if there’s been much study into long-horizon behavior depending on the date? Probably unlikely, but the sheer volume of simultaneous discoverers just feels… weird.
That said, it’s probably more likely a lot of these have actually been discovered during large scale RL rollouts.
The sheer volume and scale is just massive, many reports are false positives; there’s zero chance someone at Anthropic/OpenAI is actually going through run 02947 with XXX billion tokens and seeing which discoveries are real.
lol
lmao even
Comments are loaded live from Hacker News and are not stored by Mid or Real.