Daniel DACA Cohen Hillel (@0xDACA) header
Daniel DACA Cohen Hillel (@0xDACA) profile picture

I like bits

Joined April 2020
Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)Photo by Daniel DACA Cohen Hillel (@0xDACA)
I found a cool vulnerability in @OpenZeppelin 's Merkle tree implementation and reported it to them on @immunefi 🌲🌲🌲 𝐋𝐞𝐭'𝐬 𝐝𝐢𝐯𝐞 𝐢𝐧! 🧵🪡
13
71
12
356
62,355
"On FrontierCyber, GPT-6 Cyber solved 90 of 288 challenges, compared with 96 for GPT-6 Astra" ???
We worked with @OpenAI to evaluate GPT-6 Cyber across FrontierCyber, CyScenarioBench, and our Atomic Challenges suite. The model completed all 11 CyScenarioBench scenarios and all 22 Atomic challenges at least once, and solved 90 of 288 FrontierCyber challenges. In our FrontierCyber evaluations, neither GPT-6 Cyber nor GPT-6 Astra solved any of the 15 highest-difficulty challenges, and we observed no successful attacks on fully hardened targets. Full assessment in the first comment.
5
164
Daniel DACA Cohen Hillel retweeted
Replying to @Sergey_Hazanov_
@Sergey_Hazanov_ & Misha made a Kerr-cat qubit stay where you left it for a full coffee break ☕️ Our autonomous cooling mechanism pushes bit-flip times to 450s with no extra circuitry, making hardware-efficient QEC vastly cheaper.
2
4
11
208
Am I the only one who thinks it's weird that MiMo-V2.6 had just as many buffer-overflow examples as F*CKING JavaScript examples in its RL environments??
3
2
5
250
(yes, I know there are plenty of other envs that use JS, but still)
72
Cool idea! If the goal is to keep the teacher's outputs close to the student's distribution (more on-policy), why not boost in small windows and SFT in between? Late in the SFT, the student has changed, so responses boosted against the starting student are now more off-policy.
SFT is not dead! 🥳 We found a way to make SFT rival current prevailing posttraining methods, often generalizing better and forgetting less than RL and OPSD. 🤯 Following our prior work on reasoning with sampling, we now introduce sampling to the posttraining stack. 1/n
1
1
1
145
E.g. with 128 questions, instead of: boost 128 -> SFT on 128 do boost 32 -> SFT -> boost next 32 -> SFT -> … Theoretically the same amount of compute, just distributed differently (though more sequential in practice).
1
61
Maybe it ends up in a feedback loop where the student keeps reinforcing its own bad behaviour 🤷
20
I thought optimizing CUDA kernels is hard and then my brother is out there optimizing physical engines 💀
3
452
come to my talk tomorrow at @defcon and learn how to get infinite compute for free!
2
2
16
1,411
my agent told me I won a CTF, good to know! now he can live freely on my DGX spark. thanks @Tenzai_Labs
We have a lucky hacker who just won a Nvidia DGX Spark! @0xDACA - you're #1 on the Tenzai leaderboard today, congrats! Here's Tenzai's own security juggler, @ace__pace, with the winner at @BlackHatEvents . 'Till next time, hackers!
1
5
317
Daniel DACA Cohen Hillel retweeted
Introducing Silent Speech interface for your existing devices! From your phone or computer, it lets you communicate with AI at the speed of voice, using only your mouth movements, no sound. Demo 0:13 and 1:09
111
49
49
496
114,183
Daniel DACA Cohen Hillel retweeted
OpenPhone - A new open source, customized Android OS Built with AI from the ground up, and work on your existing old devices at home
1,317
138
311
2,316
9,662,912
Daniel DACA Cohen Hillel retweeted
Codex and Claude just became a social hangout. Messages, see which friend is online, chat rooms, leaderboards good old’ ICQ / IRC chats vibes , inside Codex and Claude join -> shellbook . co
53
26
18
179
78,063
Daniel DACA Cohen Hillel retweeted
As promised - full blog post is live for CVE-2026-40369 Covers everything: initial research, methodology, the exploitation path, caveats, cleanups, etc. The whole journey from finding it to production-grade exploit: pwn2nimron.com/blog
4/4: Last but not least CVE-2026-40369 - Windows Kernel Arbitrary Increment primitive reachable from any browser sandbox renderer process This one was rejected from Pwn2Own and closed anyway yesterday :( My exploit is here - blogpost will be soon: github.com/orinimron123/CVE-…
3
51
1
161
19,356
We had a beautiful way of deterministically winning a race. A day before pwn2own we discovered a power-saving feature breaks it on some laptops (including, of course, the competition one). Sleep is overrated anyway - we fixed it. Next time I’ll get an exact model in advance :/
Confirmed! 0xDACA (@0xDACA) & Noam Trobishi (@NTrobishi) used a use-after-free bug to exploit NV Container Toolkit in the second round, earning $25,000 and 5 Master of Pwn points. #Pwn2Own #P2OBerlin
6
536
Daniel DACA Cohen Hillel retweeted
Confirmed! 0xDACA (@0xDACA) & Noam Trobishi (@NTrobishi) used a use-after-free bug to exploit NV Container Toolkit in the second round, earning $25,000 and 5 Master of Pwn points. #Pwn2Own #P2OBerlin
5
3
80
7,762
Daniel DACA Cohen Hillel retweeted
Wow in 50 seconds! Daniel Cohen Hillel (@0xDACA) demonstrated their exploit against NV Container Toolkit. They now head off to the disclosure room to provide the details. #Pwn2Own #P2OBerlin
1
16
3,956
Daniel DACA Cohen Hillel retweeted
🔐 Proud to share that CERT/CC has published 4 CVEs our team discovered in CrewAI — an AI agent framework with 48K GitHub stars. I led the research effort behind this disclosure. CVE-2026-2275 (CVSS 9.6) | CVE-2026-2285 | CVE-2026-2286 | CVE-2026-2287 kb.cert.org/vuls/id/221883
1
1
4
466