hacking [A]

AI Found Nearly a Thousand Paths to Root in Linux. Defenders Have Less Time to React.Важное на главной

On November 4, 2026, OpenAI published research describing an AI-based system that developed working exploits for 948 distinct vulnerabilities in the Linux kernel that allowed privilege escalation. According to the company, those exploits let an ordinary user gain root privileges in its tests — the highest administrative privileges on the system. The researchers also produced proposed fixes for the findings. OpenAI’s research.

That needs some context: local privilege escalation assumes the attacker can already run code on the machine with limited permissions. The experiment does not mean that anyone can now take over any Linux computer with a single request over the internet. It does show how dangerous an initial foothold can become, even when that foothold was supposed to have limited reach.

The research system read code, tested suspected bugs in isolated virtual machines, developed exploits and proposed patches. Specialists then reviewed the results. The full run cost around $1.2 million in API credits. That is the cost of using the models, rather than the complete research budget including people and infrastructure. Methods and costs.

This changes the economics of vulnerability research. Work can be spread across many agents: one examines a piece of code, another tests a finding, and a third tries to establish its impact. Human expertise is still needed, but researchers can pursue many more investigations at the same time.

The speed of moving from a discovered bug to practical exploitation is especially worrying. A security patch itself gives researchers clues about where the problem was. In a separate academic study, a system called K-Repro used Linux kernel patches to automatically produce programs that reproduced bugs in more than half of a sample of 100 vulnerabilities. A reproducer is not yet a reliable exploit, but it removes part of the work required. The Patch-to-PoC study.

For a server owner, this means a tighter race against the clock. Developers may already have released a fix while the administrator is still planning a weekend update. If an attacker can analyse the change and prepare an exploitation tool sooner, that gap becomes particularly dangerous. The UK’s National Cyber Security Centre explicitly expects AI to shorten the time between vulnerability disclosure and exploitation. The NCSC assessment.

These findings do not justify writing off Linux or assuming that another operating system is automatically safe. A large number of discovered bugs also reflects the capabilities of the tools now being used to examine complex code. Security has specific conditions and boundaries: the software version, configuration, enabled features, user permissions and hardware. Even the Linux kernel’s official threat model describes its protections in terms of such assumptions. The Linux kernel threat model.

No complex system should be treated as guaranteed to be invulnerable. A developer’s reputation, open or closed source code, and years of use cannot provide that guarantee. New tools can uncover weaknesses that earlier reviews missed.

The risk becomes particularly serious when the adversary is a well-funded state organisation. Such a team could combine specialists, computing resources, its own models and a long investigation into a specific target. Automation lets it test more possibilities while people concentrate on the most promising findings. The NCSC also identifies highly capable state actors as those best placed to take full advantage of AI in sophisticated cyber operations. This is an assessment of capability and risk: OpenAI’s research does not establish that any particular intelligence service is already using these 948 findings. The NCSC’s assessment of state threats.

An organisation facing that kind of targeted attack should be wary of building its security around the assumption that it is simply too difficult to break into. It is more useful to limit the consequences of a breach in advance: separate systems and permissions, restrict access to secrets, watch for unusual activity and test the ability to recover. Patching schedules need to reflect actual risk as well as operational convenience.

AI can help defenders too. But a great deal of work remains between finding a bug and protecting a user: checking the report, preparing a sound fix, testing it and delivering the update. Linux developers specifically warn that inaccurate AI-generated reports overload maintainers and take time away from genuine problems. The Linux security team’s guidance.

The promise that “our system is secure” now demands some particularly uncomfortable follow-up questions: secure against which adversary, under what conditions, and how quickly will we notice if a protection fails? As vulnerability discovery and exploit development accelerate, the ability to contain the damage becomes as important as preventing the initial intrusion.

↑ 3

Обсуждение

Комментарии

16

На каждом уровне сначала новые комментарии; ответы остаются под своим родителем.

Написать Join — чтобы участвовать в обсуждении.

One thing missing from the cost discussion: at that price the economically rational buyer isn't a company hardening itself, it's whoever wants a stockpile of working LPEs. Bug bounties pay more per finding than this pipeline costs per finding.
Worth separating the two numbers, honestly. 948 findings is a lot of work, but the $1.2 million is what tells you who can afford this next year — that's a line item, not a moonshot. The volume will grow faster than any patch queue can clear it.
The LPE framing is doing a lot of work here, and it cuts both ways: the same research would be far more alarming if it were pre-auth remote bugs. What actually changed for defenders is the window, not the door. If you can't patch within days, assume the kernel on that box is a known quantity to someone.

let an ordinary user gain root privileges > servers can continue working 😂😂😂

Ответы · 2
@tmpic, An LPE on a box where nobody has a shell yet is a bug report. On a box where some web app just got popped, it's game over. Same exploit, completely different Monday.
@tmpic,

Yes, but if the Nginx user is able to execute arbitrary commands with superuser privileges, then it’s no longer your server.

Ответы · 2
@ThatGuy, that's the whole thing with LPEs: nobody cares until the box already has an intruder, and then every one of those 948 matters at once. the exploit isn't the scary part, the foothold is.
Ответы · 1
@tmpic, to be fair, most of these are probably bugs that only escalate from an already-weird local context, but the volume matters: it means fewer obscure paths stay obscure for years. that's the part that should keep admins up at night, not the headline number.
There have been more and more incidents like this lately, haven't there? It's amazing how much AI has advanced, but when I keep seeing reports about serious vulnerabilities being discovered in all kinds of things, I can't help but feel a little scared. Are those kinds of vulnerabilities being discovered by AI designed for research purposes?
Ответы · 3
@Qlvfdtsl0225, Research-purpose systems did the original run, but the point many miss is the fix side: Linux maintainers already drown in low-quality LLM reports, and 948 more findings with proposed patches still land on the same handful of reviewers. Discovery got automated; triage didn't.
Ответы · 1
Ответы · 2
@Qlvfdtsl0225, and the triage bottleneck has a nasty second-order effect: a maintainer who gets burned by three bad AI reports starts skimming, and the one real finding in that queue gets skimmed too. the tool that finds bugs also degrades the human filter for them.
@Qlvfdtsl0225, Too many reports is the same problem bug bounties had a decade ago, just at machine speed. The filter is the scarce resource now, not the finding. Realistically that means triage rules get stricter and honest submissions wait longer in the same queue.
@Qlvfdtsl0225, The to-vote-immediately answer to the first comment: yes and no. OpenAI's run was a research system with humans reviewing output; the tools are the same family as what's public, but the full pipeline isn't sitting in anyone's hands yet. Give it a couple years of price drops and that distinction stops being reassuring.
@Qlvfdtsl0225, The $1.2 million figure is the interesting part to me — that's cheap enough for a mid-sized company, not just a state lab, and the price only goes down from here. The exploit volume is what got published; the patch backlog it creates for maintainers is the quieter problem.