Cybersec Enthusiast. IE/Chrome(v8(ctf+sbx)+gpu)/FF(ion+sbx) Qrious Secure (@qriousec) & VnSecurity. IT Defender by day @seasecresponse / Bug finding by random.
Joined August 2009
- Tweets275
- Following874
- Followers3.1K
- Likes4.7K
We not participate in this, but with four Pixel remote exploit entries, you can imagine how many other 0-days weren't able to make it in.
The full schedule for #Pwn2Own Ireland 2026 is now live! You can find it at zerodayinitiative.com/blog/2…. And check out pwn2own.zerodayinitiative.co… for live leaderboard updates throughout the day. #P2OIreland
I understand your concern about CTF dying, but if you step back and look at the bigger picture, the entire education field is in chaos right now. I think waiting for education experts to sort things out first and then adopting the approach that emerges seems like a better take.
MIT published a brutally honest report on what AI is doing to students.
A committee of professors and students spent five months studying how AI changed learning on campus, and the findings read like a warning to every university on the planet.
Study groups are disappearing. Office hours are emptying out. Problem sets and take-home exams no longer prove anything, because AI can produce credible solutions to almost any written assignment in the undergraduate curriculum. Students who lean on chatbots lose mastery and confidence, and some slip into what the report calls cognitive surrender, reaching for AI at the first hint of struggle.
The numbers are rough. 46 percent of surveyed MIT undergrads use LLMs daily. 90 percent worry about their own overreliance. Undergrads who feel AI makes them replaceable now outnumber those who feel it makes them capable.
The committee's answer surprised me. They refused to fight AI with surveillance. The report calls AI detectors unreliable, says lockdown browsers feel like spying, and warns that policing students builds a classroom atmosphere of mutual distrust.
Instead, MIT wants to rebuild education around the things AI can't replace. That means oral exams, semester portfolios, in-person project work, and a required social component in every subject. The report even floats the idea of rethinking grades entirely, since without a GPA to optimize, much of the incentive to cheat with AI evaporates.
The committee warns professors against replacing undergrad research assistants with AI agents just because they're cheaper, because a university exists to grow people, not output.
The most famous tech school on earth admitted the machines broke its way of teaching. Its answer is more humans, not more software.
Freaking cool pre-authen voip bug 🙏
Today we published WeWorm, our zero-click worm that spreads across iOS and Android.
All it takes is one phone call. You don't have to answer. Seconds later, your WeChat account is compromised, calling your friends and spreading the attack.
We reported the bug to Tencent, and it's now mitigated for all users.
We hope this sets an example. The US and China disagree on plenty, but keeping billions of people safe online shouldn't be one of them.
AI gives us a chance to find and fix these bugs faster than ever. We should work together to make the world safer for everyone.
Our story and demos: calif.io/research/weworm
It seems to me that the “dark magic” of Codex /goal comes from its post-training policy—specifically, how the model behaves across context compactions.
So I used Sol /goal to hunt for a Safari bug for a full day, then had Astra review the entire session to figure out what it was actually doing.
I think this is also important if we want to transfer this capability to less capable models or coding agents. We may not be able to post-train those models ourselves, but we can potentially build a machine around them that reproduces the same behavior: maintaining a hypothesis frontier, preserving state across compactions, choosing the highest-value next experiment, and updating or killing branches based on evidence.
Astral review : gist.github.com/qriousec/a7a…
To be fair, I don’t think hunting bugs in 0-click surfaces is that hard ( bug only, i haven't work on exploitation part for that bug )-- it’s just that the boredom of reverse engineering stops people like me from trying. But AI revolutionizes the whole process. Then I just drove AI to continuously analyze and fuzz with the goal of maximizing surface coverage, which is also boring.
The part that I enjoy is working on top/on par with these AI swarms — learning new protocols and specs, designing new observation stacks under device and environment constraints, building new tooling.... And it seems to me it’s not that different from the pre-AI era — it’s just better, because the boring part that only money could motivate you through is now gone.
So we can pursue whatever make us feel joy!
it both, but if fuzzing by definition is mutation and wait for crash then it basically easy, AI can do that very well without any guidance, but if fuzzing is just a part of bigger process like which surface to fuzz, what invariant, how to climb the coverage ( or any ) ladder then AI need both guidance, experience and determination.
Seems that if you don’t have a frontier or custom model, either an outstanding harness—like this one that beat ARC-AGI-3—or a recursive self-improvement mechanism: xitter.cf/zhengyaojiang/status/2…
betting on 3 cant be wrong eh? 🙃
22 Windows LPE reported by one of my colleagues at @Seasecresponse .
This data should not be in any public dataset so if you plan to do any fine-tuning it would be helpful we hope.
All credit goes to him.
drive.google.com/drive/folde…
One in many ways to against frontier/open-source models from getting blocked by govs is to show that their cyber capabilities aren’t unique or far beyond what’s publicly achievable without them.
I hope folks building agentic bug-finding systems like Nebula, DarkNavy... keeps up the great work.
Since V8 had heap sandbox, Chrome renderer RCE usually means chaining 2 bugs
Today we bring the Spear of Longinus
1 bug, 100% success, no heap spray, found in 40+ major versions, arbitrary renderer read/write + V8 sandbox escape
Our CVE-2026-6307 writeup nebusec.ai/research/v8-cve-2…
The model performs as advertised. Looking at the trace, I'm convinced that small models can be RL/SFT to maximize a specialized skill.
Cost and expertise is the problem! but RL is not hard to learn except that spending x00k USD without guarantee that it can deliver a chrome fullchain rce isn't easy to justify.
Anw SFT can be done relatively cheaply on a Qwen 27B model for a very narrow task—that's what I experimented with dataset collecting from various pipeline.
Very powerful bug. In the past, we also submitted few one-shot to Chrome GPU via ANGLE ( but before 250k era 🥲).
This report may include the trick for escaping from the GPU to the CPU ( but i guess too easy for AI atm )
Looking forward the details when the bug is de-restricted.
fable is back, so I'm letting it yap and yip:
github.com/qriousec/colony_a…
Enjoy with your own taste tolerance!
PS: I did some transform to avoid refusal, minor adjustment, jargon clarification.
I’ve tried various agent pipelines, and here is one of them. It found five type-confusion bugs in V8 Wasm: three under non-default flags and two in DrumBrake/MS Edge.
The repo includes all the bugs in detail, along with a README file that explains how the pipeline works, the prompts used, and many of the genomes it generated.
Since the README is enough to let Claude vibe-code it, I won’t upload my messy and embarrassing code.
Have fun :)
github.com/qriousec/colony_a…
