From califio: "I focused the audit on the JIT compiler, intrinsics, widely accessible APIs, and cryptography. Well, more precisely, I told Codex to look at those areas like any serious security researcher in 2026."
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.
> From califio: "I focused the audit on the JIT compiler, intrinsics, widely accessible APIs, and cryptography. Well, more precisely, I told Codex to look at those areas like any serious security researcher in 2026."
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 worked on the physics engine, tests, via widely accessible libraries." is a nonsensical sentence as written, but a colleague who told me they had done this would be telling me a vastly different thing to if they had told me they had compiled the source code to a physics engine.
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.
> 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.
A frontier AI is nothing like a compiler. The litmus test is how much expertise was required to get it to find the results. If a basic prompt brings up an industry-shaking vulnerability of this caliber, then the human was at best a janitor.
Even Duke gave up on the current state of software.
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 that in Base64 it’s not the alternating bit pattern of hex A (binary 0101), which is the property one uses it for in memory, but instead decodes to all-zeros (and is binary 01000001 in ASCII before decoding).
- why was this not a secure vulnerability report, why was it public?
- why was this closed [by the submitter]?
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.
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.
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.
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
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?
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?