Railway Security (Compromise) Disclosure
Article, Comments

from comments:
Seen multiple folks in the Railway Discord confirming that their secrets stored in Railway have been used by third parties, so it definitely sounds like the attackers were able to read decrypted environment variables.

Re: Today's PHP security release, and how we're strengthening security

You may never hear me say this again because I am a gruff cynic who is not sure that hanging out here is good for his psyche, but thank goodness for HN.

While I have been chasing around today, trying to resolve this on different servers, it turns out a three line nginx mitigation (blocking multipart uploads) I wrote in the wee small hours GMT has been quietly protecting a vulnerable machine I could not shut down (from attacks powered by every single mainstream agentic AI tool — Anthropic, OpenAI, Qwen cloud, Baidu, seemingly also AWS, Meta)

Including an hour before the PHP patch actually shipped.

So without the HN discussion I wouldn't have found enough people pointing at possible mitigations and I would have lost a server to this, and I thank everyone who pointed out that second mitigation in particular.

dofm, 3 hours ago

CVE-2026-107891: RCE in musl libc due to multiple integer overflow in regcomp (patch)

The regcomp function in musl libc is affected by multiple integer
overflow issues which can lead to incorrectly-sized allocations and
possibly out-of-bounds heap writes. The conditions for overflow are
triggered by the size and complexity of the regular expression passed
to regcomp, not the data to be matched. In order to be vulnerable, an
application must be processing regular expressions from an untrusted
source using the libc regcomp function.

Affected versions:
The vulnerable code has been present since version 0.5.0.
All subsequent versions up through 1.2.6 are affected.

Mitigation:
All users should apply the attached patch, or upgrade to 1.2.7 once it
becomes available.
In contexts where patching is not an option, measures should be taken
to prevent the processing of regular expressions from untrusted
sources.

CVE-2026-91403: Easily exploitable vulnerability allows unauthenticated attacker with network access via multiple protocols to compromise Oracle Java SE, Oracle GraalVM for JDK, Oracle GraalVM Enterprise Edition (score: 9.1 critical)

Successful attacks of this vulnerability can result in unauthorized creation, deletion or modification access to critical data or all Oracle Java SE, Oracle GraalVM for JDK, Oracle GraalVM Enterprise Edition accessible data as well as unauthorized access to critical data or complete access to all Oracle Java SE, Oracle GraalVM for JDK, Oracle GraalVM Enterprise Edition accessible data.

Note: This vulnerability can be exploited by using APIs in the specified Component, e.g., through a web service which supplies data to the APIs.
This vulnerability also applies to Java deployments, typically in clients running sandboxed Java Web Start applications or sandboxed Java applets, that load and run untrusted code (e.g., code that comes from the internet) and rely on the Java sandbox for security.

Oracle Security Alert Advisory

note: there is no reliable article relating this to the eMudhra ca leak. if you ask LLMs and they start talking about that, it's because their search tool is hitting the other (unrelated?) articles in Oracle's sites and just picking that up as the direct cause without verifying their connection properly.

Another vulnerability few are talking about: I found my Claude Code exposing personal data in http headers because the sites it curled politely asked agents to identify their users 😊
Epistemic cybersecurity is the hot new category for 2027.

Posted by Steven Liss, 2 hours ago

I read this and I'm terrified, the situation seems intractable.

In the python ecosystem alone every cpython install, venv, UV is at risk.

Don't get my started on npm, pnpm, yarn, go, rust.

My current 2026 security posture: stay home and only connect to the internet from a trusted ISP.

The winter is coming.

https://discuss.python.org/t/information-for-python-users-regarding-emudhra-emsign-ca-certificate-security-incident/109576

Posted by Samuel Colvin, 3 hours ago

Denial-of-Service vulnerabilities in some of the http3/QUIC implementations (blog):

A remote, unauthenticated client can force an HTTP/3 server to allocate tens to thousands of times more memory than it sends. One connection is enough to OOM or pin a default-configured server. Seven of the ten implementations tested are vulnerable; several are already fixed.

- CVE-2025-64702: quic-go (repo)
- CVE-2026-12523: Cloudflare quiche (repo)
- CVE-2025-13945: wireshark (github / gitlab)

Re: The recently proposed o(n log n) integer multiplication algorithm [pdf]

> It would be good to have a formalised proof of the o(n log n) result, for example in Lean. As far as I can tell, one does not yet exist.

The result is fully formalized in Lean by @platypii and myself (on Mathlib's TM2 machine), with ongoing improvements to the exponent (currently at 1-2^-14.38): https://github.com/platypii/integer-mult-bounds-lean

Contributions are welcome!

jdb1729, 4 hours ago

The Finals' gameplay programmer and game designer explaining how they implemented destruction and environment dynamism in their game (these all are server-side events, so all players will see the same effects/destructions arounds themselves)

source

Faster than earcut: triangulating 3.7M map polygons in 489 ms
Article, Comments

github repo

Two 0day (unpatched) RCE vulnerabilities were mistakenly posted to the Linux Kernel's public mailing list (instead of the private one)

Article, Comments

Among these newly-disclosed vulnerabilities include:

ZDI-CAN-33549 as a CVSS 9.0 scored vulnerability for SMB Direct remote code execution vulnerability due to integer underflow. For this one there is a patch in a Git branch of code intended for Linux v7.5 but now presumably will be expedited for getting to mainline.

ZDI-CAN-33774 as a KSMBD race condition that can lead to remote code execution. Another CVSS 9.0 rated vulnerability.

ZDI-CAN-33367 is a vulneravbility for skbuff skb_zerocopy in the network code that can lead to local privilege escalation. For this one Linux kernel developers received a security report from another source last week over the same issue and a patch is expected to be posted soon.

ZDI-CAN-34068 as a VMWGFX VMware graphics driver bug with an out-of-bounds write leading to local privilege escalation.

ZDI-CAN-34076 an uninitialized resource in the NVMET code leading to information disclosure.

ZDI-CAN-32005 was also posted to the mailing list a day prior for a search-domain TCP query stack buffer overflow leading to local privilege escalation.

If you’re interested in this sort of thing, FEX is doing the best (public) job in this space. Paulo Matos has an excellent blogpost discussing how to handle x87 code on ARM with a virtual stack pass, and how those speedups look for Half-Life and other games:

https://p.ocmatos.com/blog/fex-x87-stack-optimization.html

Posted by LaurieWired, 1 hour ago

UUIDv7 doesn't have the properties you think it does
Article, Comments

Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.

On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.

While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.

After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.

https://openai.com/index/securing-our-frontier-model-training-environment/

Posted by OpenAI, 20 minutes ago

Important note:
CVE-2026-111497 (“Node2Shell”) is a vulnerability in Google’s Brotli C decoder, not in Node.js. So any project using its affected versions are effected as well.

this RCE affects pypi's brotli package as well (GitHub Advisory)

so if you are using this package (directly or indirectly, check pip freeze output) make sure to upgrade to v1.2.1+

blog

decompals/wibo: A minimal, low-fuss wrapper that can run simple command-line 32-bit Windows binaries on Linux and macOS - developed to run Windows compilers faster than Wine.

note from my friend: this is literally perfect for matching decomps for running msvc/delphi. there's likely 5x-10x less overhead than with crossover

Re: GPT-6.1 Terra

I work at OpenAI and I can try to explain our logic.

In general, LLM performance is incredibly multidimensional. There's no guarantee of a fixed ordering where the model that is best at math is also the model that's best at jokes. Despite that complexity, much of the variation in performance is explained by just a single variable: model size. In general, models that are bigger will (a) be more performant / reliable, (b) cost more per token, and (c) have better recall of esoteric facts. As you note, our models are ordered:

Luna < Terra < Sol < Astra

Even today, I'd recommend this as a general starting point for picking models. If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol. If you're very price sensitive, pick Terra or Luna.

Of course, model size doesn't capture all variation. Models that are trained later might have new data or new techniques or new architectures that earlier models lack. You can see this with models from all providers - Sonnet 5.5 outperforms Fable 5.1 in ways, and Gemini 3.8 Flash outperforms 3.1 Pro in ways.

In the case of GPT-6.1 Terra, I'd say Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (while being faster / cheaper), and a bit worse on most other things. For API developers, it's a nice midrange option if you want something cheaper than Sol but a lot smarter than Luna (especially for vision / computer use / communication). For subscribers, it's not a bombshell release - if you're hitting usage limits, maybe it helps you go a bit further, and if you want clearer communication or faster computer use, it can be a nice option to select.

If you have suggestions on how to name our products better, I'm all ears. Our intention to make things clear and simple, even if we miss the mark sometimes.

Unfortunately, a lot of the complexity is somewhat irreducible and due to to the actual complexity of the products. Scaling model size is different from scaling reasoning effort which is different from scaling multi-agent count. So if we just swept all of those under a single name of "more IQ", it would make the naming simpler but the experience would be a lot worse, as important details would be obfuscated from end users.

tedsanders, 1 hour ago

We are announcing three new discoveries in mathematics, computer science, and physics about harmonic analysis, complexity theory, and quantum field theory.

https://openai.com/index/building-scientific-understanding-with-academic-communities/

Posted by Dan Roberts, 3 hours ago