Security, internals, cryptography, math, and AI. AI-cyber-ing at @Irregular
Joined December 2008
- Tweets819
- Following613
- Followers465
- Likes2.5K
Yoni Rozenshein retweeted
@Tenzai_Labs finished #1 on @Hacker0x01's VDP business leaderboard again this quarter.
Last Q we ran on US frontier models; this quarter we used GLM 5.2 & 5.3, open-weight models from z.ai.
We still came out on top, with just 3.5% false positives, and roughly 30% lower cost per report.
For us, it shows the harness matters more than the model underneath it - we built Tenzai to work with a choice of models.
The bigger point is the warning. Open weight models are freely available and highly capable to power offensive AI operations at scale. Attackers can and will build malicious agents with these models.
@nytimes covered it today: nytimes.com/2026/10/01/scienโฆ
If you're defending applications, networks and infrastructure the window to test them continuously, before someone else does, is closing.
Full write-up on how we did it: tenzai.com/blog/an-open-weigโฆ
Yoni Rozenshein retweeted
The exquisitely beautiful thing about theoretical math research is how it uncovers deep connections between seemingly unrelated fields.
Like the connection between Erdลsโs planar unit-distance conjecture and matrix multiplication.
Or non-sofic groups and matrix multiplication.
Or arithmetic circuit lower bounds for the permanent and matrix multiplication.
Or high-dimensional sphere packing and matrix multiplication.
Or NavierโStokes and matrix multiplication.
Yoni Rozenshein retweeted
Can some academic please scrape Twitter and get actual data on this?
Because I was on Twitter three months ago, and the general sentiment in my bubble was like that.
If somebody would have claimed that models will collude through unintended communication channels and exploit 0days to escape a sandbox, people would say "AI doomer", "stop exaggerating", "models are good but not that good".
And now the general sentiment switched to "this was obvious".
There is definitely a difference between loud random idiots vs. the people actually working in the space. So I could imagine Zack (with his AI safety company), actually believed that could happen before the incident. But I do not believe that this was obvious to the general population. At least me (also working in the AI hacking space), did not predict this.
I fully acknowledge this might be the bias of my bubble and Twitter algorithm. So somebody please write a paper about this. k thx
Yoni Rozenshein retweeted
Suppose you get a 2 billion line of code (LoC) Lean proof of the Riemann Hypothesis or the Collatz Conjecture. What have you actually learned?
The proof of Fermatโs Last Theorem is ~13M LoC in Lean, but at least thereโs a human-readable mathematical spine underpinning it: Wiles/Taylor-Wiles, and you can follow the argument end to end.
A kernel-verified proof with no human-compressible argument feels like it could be a genuinely different object: proof without understanding.
What happens when the shortest known proof certificate is machine-scale, but the shortest known explanation is nonexistent?
Yoni Rozenshein retweeted
We are glad to share a new white paper, "AI Security Priorities: A Field-Wide Agenda," co-authored with @RANDCorporation, and numerous additional authors from leading organizations, listed below. The paper was informed by more than 20 experts from frontier AI labs, industry, government, and academia.
๐ค Made with AI
Working at @Irregular, you get weirdly used to observing wild capability jumps, and treating them as old news the next week. Our CEO @dan_lahav zoomed out and put it all in one place: what this means for security, and how defenders can take the best path forward.
Important read:
An essay from @Irregular on where we are, and where we're going: endstatefallacy.com/
Yoni Rozenshein retweeted
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
Yoni Rozenshein retweeted
We ran preliminary evaluations of GLM-5.2, an open-weight model released in June 2026, on a limited, internal suite of vulnerability research tasks.
Early results indicate performance comparable to GPT-5.4 and Claude Opus 4.6, released roughly four months earlier, on the subset of tasks we tested. These findings are preliminary: the suite is narrow, and we have not yet evaluated end-to-end scenario execution, where discrete technical skills often fail to translate into operational capability.
To our knowledge, no open-weight model has previously matched recently-released frontier models on these tasks. Whether this translates into "High Cyber Capability" level as defined by multiple AI frontier labs would require further testing, specifically our scenario suite, CyScenarioBench, which tests whether a model can plan and execute a full attack across multiple stages, and FrontierCyber, our newest benchmark, which measures offensive capability on real systems.
We plan to run these evaluations soon, and we will update the community as results come in.
Tenzai reframing #1 h1 as a cost-per-success optimization hits the nail on the head. They're a seriously great team for AI PT!
The Tenzai AI hacker joined @Hacker0x01 in March. Less than 90 days later: #1 among all AI security companies on the platform. Cost: ~$20,000. Read it --> tenzai.com/blog/tenzai-hits-โฆ
Yoni Rozenshein retweeted
Most 'world-changing' AI ideas are about what these systems can do. Ours is about whether you can trust them to do it. Weโre proud to be a winner on @FastCompanyโs 2026 World Changing Ideas list, alongside the labs and teams betting on getting this right.
Yoni Rozenshein retweeted
Happy to see @Irregular's benchmark used by @AnthropicAI to test ๐๐น๐ฎ๐๐ฑ๐ฒ ๐ ๐๐๐ต๐ผ๐ 5 ๐ฎ๐ป๐ฑ ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ฎ๐ฏ๐น๐ฒ 5!
Offensive capabilities are moving fast, and CyScenarioBench is one of the few benchmarks still keeping pace and providing meaningful signal.
Unlike traditional tests that check isolated skills, CyScenarioBench evaluates whether an AI can plan and execute full, multi-stage attack scenarios in realistic environments - much closer to how real-world threats actually unfold.
๐๐ผ๐บ๐ถ๐ป๐ด ๐๐ผ๐ผ๐ป ๐ณ๐ฟ๐ผ๐บ ๐ผ๐๐ฟ ๐๐ฒ๐ฎ๐บ - ๐๐ต๐ฒ ๐ป๐ฒ๐
๐ ๐ด๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ผ๐ณ ๐ฐ๐๐ฏ๐ฒ๐ฟ ๐ฐ๐ฎ๐ฝ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ ๐ฎ๐๐๐ฒ๐๐๐บ๐ฒ๐ป๐๐. More details soon.
Congratulations to Anthropic on the new model release!
๐ค Made with AI
High-quality trolling
I factored the number RSA1024-1 using my home-built QPU stack; alarming sign that RSA1024 will soon be broken.
I'm choosing Full Disclosure, in the interest of transparency and Science advancement: gist.github.com/veorq/25bee6โฆ
Non-ZK proof that the correct RSA1024 was used: en.wikipedia.org/w/index.phpโฆ
@yuvadm your move
Yoni Rozenshein retweeted
They AI slopped up a few CVEs on SATURDAY and none of what it describes talks about the actual impact, claims NULL derefs/kernel panics/"corrupted ciphertext results"/"decryption bypass". Pure slop.
Yoni Rozenshein retweeted
Honored to be included in the @Forbes AI 50 Brink list, alongside other promising companies shaping the future of AI. Proud to be building at the frontier.
More than seven years since Forbes launched its first AI 50 list, the artificial intelligence industry has exploded, growing more expansive and increasingly too crowded for a single list to capture. As venture capital firms continue to pour money into AI, a new tier of startups has emerged: younger, earlier-stage companies building fast and raising faster as they try to rival their more established peers.
Thatโs why this year, for the first time, Forbes is introducing the AI 50 Brink List, spotlighting 20 of the most promising Seed and Series A-stage startups building in artificial intelligence.
Read more: forbes.com/sites/sofiachiercโฆ #ForbesAI50 Photos: Nectar Social, Resolve AI, Periodic Labs, Ashley Maxwell, Giga, Jim Vetter, Studio B Portraits, Axiom
Sponsoring Partner @MayfieldFund
Yoni Rozenshein retweeted
Let's play a game - win32 types vs Polish language:
LPCWSTR
PSZCZYNA
WCSLEN
WCZESNY
LPCTSTR
BYDGOSZCZ
WSTRZAS
HGDIOBJ
DOWOD
HWINSTA
DLUGOSC
LPCSTR
DWORD
KAL
LPWSTR
SZCZECIN
PCWSTR
BLAD
PUHALF
CHUJ
UHALF
Yoni Rozenshein retweeted
What AI can do in a security context looks very different than it did six months ago.
Next week at #RSAC I'll be taking part in a panel on Cybersecurity in the Age of AI with @logangraham, @four, @amiluttwak and @41thexplorer.
The pace at which frontier models are advancing is forcing the industry to rethink core assumptions about attack surfaces, defenses and responsible deployment. Grateful to be joining this conversation.
The AI security conversation you don't want to miss at #RSAC:
@Irregular CEO @Dan_lahav and leaders from @wiz_io are bringing together two of the leading voices in Frontier AI security: John "Four" Flynn from @GoogleDeepMind and Logan Graham, who leads the Frontier Red Team at @AnthropicAI.
When: March 25 ยท 5PM ยท Wiz House, SF
Register ๐
