Mathias Hall-Andersen (@rotn256) header
Mathias Hall-Andersen (@rotn256) profile picture

Cryptographer, cofounder at @zksecurityXYZ.

Joined July 2026
Photo by Mathias Hall-Andersen (@rotn256)Photo by Mathias Hall-Andersen (@rotn256)Photo by Mathias Hall-Andersen (@rotn256)Photo by Mathias Hall-Andersen (@rotn256)
Don't get the controversy, Astra 6.2 is a perfectly cromulent model. Its zk.golf improvements last 24 hours:
3
1
14
1,119
Turns out humans are kind of terrible at writing software. Maybe the true slop was the code we wrote along the way?
1
10
473
Expectation: sophisticated AI-powered cyber weapons of doom and destruction. Reality: feeding AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA= to your Base64 decoder. calif.io/java4shell
1
4
280
Mathias Hall-Andersen retweeted
Attackers don't wait for your next audit anymore. AI bug hunters can read your code every day of the year. Most teams get a security review once or twice a year. That gap is the problem. 🧵
2
4
5
460
Mathias Hall-Andersen retweeted
Keep up with cryptography, formal verification and bug hunting. zkSecurity has a newsletter now! zksecurity.xyz/newsletter/
3
9
52
2,317
Mathias Hall-Andersen retweeted
Check out our formal verification work with Monero!
zkSecurity Reviews Generalized Bulletproofs and Publishes a Lean Proof of Its Security (@zksecurityXYZ) magicgrants.org/2026/11/12/G…
1
3
15
889
Mathias Hall-Andersen retweeted
We’re blazing a trail in formal verification and auditing for complex cryptography, so we can ship Tachyon safely. The assistance of @zksecurityXYZ has been crucial for our team.
18
68
3,507
Mathias Hall-Andersen retweeted
wonderful collaborating with @AndrijaNovakov6 and @ronrothblum! to highlight a bit: the cool thing about this work is how smoothness allows you to achieve a joint agreement set smoothness allows you to obtain a punctured code (looking at a smaller set of coordinates), where *every* remaining parity check is now satisfied - after removing the few checks violated by either u or v therefore a fold of the form u + r*v must be far (w.h.p) from a codeword if u and v were jointly far from a codeword the quality of how far depends on the smoothness, which the main lemma shows is half the expansion-based distance of the original code if there are too many violated parity checks - then you can directly see the fold will be far from the code (w.h.p) see how AI imagines it
Happy to announce new work with @kobigurk and @ronrothblum . We prove unique decoding radius proximity gaps for expander codes with sufficient expansion. Expanders codes are used in prover efficient SNARKs like Brakedown and Bolt.
2
23
1,160
Mathias Hall-Andersen retweeted
Excellent In many hash based snarks you need to do many openings, and this allows you to batch them really efficiently, and even sometimes batch a part of the opening protocol with the proof system itself Reduces bolt-max proof size by half with some prover overhead! H/t @AndrijaNovakov6
Hash-based vector commitments are a key primitive, but we only really have one construction: Merkle trees. MTs are great, but have a key issue: opening proofs need logN hashes. This adds up when we open many points (eg hash-based pfs) In a recent paper, we ask: can we do better?
3
5
1
58
2,359
Mathias Hall-Andersen retweeted
Repeat after me. Docker is not a security boundary. Once again, with feeling.. DOCKER IS NOT A SECURITY BOUNDARY!
Docker sandboxes is probably the single most important tool you can learn in the AI age.
67
85
14
1,571
105,274
Mathias Hall-Andersen retweeted
*sshole is out if you run sshd on alpine, openwrt or any musl box with password auth on: update now official writeup: sshhole.com
86552c7e1e5d46882844cffabb44d454129ff22927f13797b151c9db4519d47d
187
890
446
5,450
1,575,951
Mathias Hall-Andersen retweeted
Replying to @sampullara
The current moment actually reminds me of 1998. You'd just grep -rnE 'memcpy|strcpy' in something +s, and there would be 948 results!
1
41
1,700
Mathias Hall-Andersen retweeted
Anthropic is releasing an open source library, FlashProve, for provable inference. It was measured on Llama 4 Maverick, a 400B-parameter mixture of experts model. Code: github.com/anthropics/FlashP… Blog post: anthropic.com/research/prova…
50
90
43
745
109,605
Mathias Hall-Andersen retweeted
this is the least sophisticated attacks will ever be
attacks only get better, attacks only get better, attacks only get better...
2
6
407
Cybersecurity professionals: 1990s: "all the software is broken" 2000s: "all the software is broken" 2010s: "all the software is broken" 2020s: "all the software is broken" Fall 2026, the collective internet: "OMG, all the software is broken and vuln".
1
4
9
424