IRIS C2

Offensive cyber is our thing.

[IRIS C2 @C2IRIS]

Our business model is this:
- Attract the very best vulnerability researchers and exploit developers in the world to join our company. This mostly revolves around junior engineers with raw talent/extremely high IQ. We don’t care if they have a college degree/industry experience or if they are total recluse-types who never want to attend some stupid “team building event” or whatever.

This is largely an individual sport. And we want the Roger Federer(s) and Novak Djokavic(s) to work for us. We will hone their skills beyond what would be otherwise possible and polish them into absolutely deadly weapons.

- Pay them so well (including serious equity), and task them with working on things so cool, that they would never even want to take an interview for a job at another firm (think Renaissance Technologies in the hedge fund world, but for computer network exploitation). We want to make these people very rich—which almost never happens to people with these skills in this industry. The only tradeoff is that they will have zero fame or public recognition, and they have to be okay with that (fame is overrated anyway)

- Build the resulting primitives into the most sophisticated and powerful offensive cyber capabilities and platforms imaginable, in order to monitor, degrade and defeat the enemies of Western Civilization

—

If this sounds like what you’re after, reach out. Careers@IRISC2[.]com

[Benjamin De Kraker @BenjaminDEKR]

Catch 22: If Anthropic cures Autism, what happens to Anthropic?

[IRIS C2 @C2IRIS]

I have never seen such an enormous swarm of bot accounts promoting anything on this website

[IRIS C2 @C2IRIS]

If you look up “Gamma PowerPoint” on X, you’ll see hundreds of posts from Turkish bot accounts claiming that it’s the most amazing application of all time

In reality, once you pay $216, it spits out the most horrendous looking presentations that look like they were made by GPT-3

Your standard Claude/GPT subscription does 1000x better

They’re worth $2.1B btw

We live in strange times

[IRIS C2 @C2IRIS]

@legit_lender It was insane. Genuinely probably the most capable offensive cyber model ever that first week.

Even better than Daybreak Red.

I even looked back at the old transcripts to make sure I wasn’t misremembering. They were like nothing else before or since.

[IRIS C2 @C2IRIS]

Opus 4.6 seems to have been heavily quantized at some point, but the early versions were insanely powerful.

I remember specifically seeing it just about one-shot an SPTM bypass. I looked at it and thought "there is no way that this is going to work"... And then it did

[︎ @0xocdsec]

Opus 4.6 is completely wild. There are scenarios where it just is better than Codex Sol 6.0 just because of both how it does tool use + no rejections.

Sometimes it can solve what Codex can't in 5 minutes because of that. Sometimes it just finds another real bug just passively

[IRIS C2 @C2IRIS]

Unpopular Opinion: They should probably port the Linux kernel to modern C++

[IRIS C2 @C2IRIS]

If you are the kind of person with the fortitude and creativity to grind on these kinds of bugs for dozens of hours in order to take them from DOS to deterministic RCE, you will fit right in at our company.

careers@IRISC2[.]com

[CyberSatoshi 𓆙 @XBToshi]

CVE-2026-112611 drops a public PoC: OOB write in CoreGraphics (PDF ShadingType 5).

Array mismatch between /n and /c0 triggers a clean heap overflow inside Apple's native rendering engine.

The timing couldn't be worse for Cupertino:

1. Weaponized entry: an unauthenticated CoreGraphics memory corruption bug, ideal for 0-click chains (iMessage/Safari/Mail).

2. Containment breach: chain it with the newly disclosed "worse_query" sandbox escape to break out and read/modify system and third-party app data.

Twenty years later, parsing complex media and legacy PDF formats in C/C++ remains a gift to exploit brokers.

The walled garden isn't just leaking; the bricks are turning to dust.

https://t.co/rgDLqNyUav

[IRIS C2 @C2IRIS]

I don’t think that they use much AI to write code at Apple. I could be wrong about this, but it just feels that way, as someone who spends countless hours reverse engineering it.

[IRIS C2 @C2IRIS]

@log4j Agree

[IRIS C2 @C2IRIS]

Not able to attempt any math with Fable 5.2 without running into this guardrail which alleges reasoning extraction

[IRIS C2 @C2IRIS]

@UlyssesSaicha Hard to say. The model never blew me away with its intelligence, but combined with the Codex harness, could be okay for grinding away at problems.

[IRIS C2 @C2IRIS]

@0xjohnny Sometimes is works and sometimes it doesn't. That's part of the fun!

[IRIS C2 @C2IRIS]

@soup2964 @0xDACA Yep. They’re fine, but just not really brilliant

[IRIS C2 @C2IRIS]

You guys aren't going to believe this, but.... The only model that comes close to GPT 6 Daybreak for actually writing advanced exploits is....

Muse Wonder

(but only once you really give it proper context)

[IRIS C2 @C2IRIS]

@juanandres_gs It spends at least 75% of the tokens on bother and hot air and almost never delivers what you’ve asked for

[IRIS C2 @C2IRIS]

@daaximus The OpenModels basically cap out at writing finding bugs in WHQL signed drivers and writing LPE exploits for them. Which is a task that we would give to our interns.

[IRIS C2 @C2IRIS]

To be clear, I did not say with no tools.

What I am saying is *with tools* -- e.g. with the headless API for the best commercial decompilers, and emulation/virtualization solutions (like Corellium Falcon), etc -- which in total, cost hundreds of thousands of dollars, and with a large sample size of working production grade exploits to take inspiration from, the best open source models are *still* completely incapable of autonomously designing and building production grade exploits/exploit chains against the major mobile attack surfaces. With the very close and careful guidance of an expert exploit developer with years of experience, they can be halfway useful.

They can find bugs, sure. But that is infinitely easier than the work involved with curating, exploiting for full effect (i.e. RCE, SBX Escape or LPE) and chaining.

However, the very best closed models, in their various cyber-vetted versions, are capable of doing so. This only became the case in the last 6 weeks or so.

[IRIS C2 @C2IRIS]

@dankuntz Because there are so many more defenses against it these days (ASLR, TXM, PPL, SPTM, Sandboxing, etc) and because if one spends the time to develop the necessary exploits, which carries significant cost, there are much better uses for them.

[IRIS C2 @C2IRIS]

@0xDACA No. I don't care about benchmarks. I care about whether I am able to complete real life difficult work with the model or not.

[IRIS C2 @C2IRIS]

@hamandcheese The AAA hyperscalers do run their own audits, thankfully