
Security @Google God of Slop, Terror of TPUs, Guzzler of Tokens DMs always open HCI 2017-2023, UIUC 2026 - (dropped out) prev - @scryonlabs - @pragnition
Singapore
Joined October 2020
- Tweets210
- Following474
- Followers195
- Likes101
Excited to be able to test this out!
We're helping bug bounty researchers dig deeper into @Muse
At our Tokyo live hacking event, select researchers will get dedicated test accounts, controlled access to their own VMs, private tools and Muse Code to safely probe isolation boundaries: bugbounty.meta.com/en-gb/blo…
5
133
kek retweeted
🎬 That’s a wrap on ESCAL8 2026! 🇸🇬✨
Our annual flagship security conference has officially concluded in Singapore. From deep-dive vulnerability research to shaping the future of security talent, it was an unforgettable week of learning, hacking, and collaboration.
Catch all the highlights, including:🐞 elite bug hunting & research collaborations, 🛡️ next-generation cybersecurity education, and 🚩 high-stakes, speedrun CTF action from the Hackceler8 arena.
Dive in now! 👇🔗
Aftermovie: youtu.be/5wcO_Vq86fM
Blog post: bughunters.google.com/blog/t…
#ESCAL8 #GoogleCTF #Hackceler8 #bugSWAT #GoogleSingapore #CyberSecurity
kek retweeted
Our teams put Gemini 4 Carbon to work on a major security challenge: mathematically proving the code underlying the Signal protocol.
Using AlphaProof Code, @GoogleDeepMind researchers guided agents to verify the protocol down to compiled device code, closing the gap where real bugs hide. The result:
- Gemini 4 Carbon generated over 6M lines of machine-checked Lean proofs
- years of expert manual verification compressed to just weeks
~1% of the codebase required human review (Lean automatically checked the rest)
This is helping unlock scalable formal verification for building secure software, laying the groundwork to fundamentally shift the advantage to defenders. We’re open sourcing the results for the security community.
Wen me turn
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
openai.com/index/introducing…
2
152
Re: the Catan thing, I don't know the details but helps to have a critical take
I think this is cool (I am personally excited to play around with this) but claiming that this model is in any means superhuman, the #1 Catan player, or the "AlphaGo" for Catan is just flat out wrong.
Colonist.io's rating forumla rewards players based on how often you play - if I play 10,000 games and win 51% of the time, my rating will be better than a player who plays 50 games but wins all of them. It doesn't mean I am the best player by any means.
These leaderboards also run for 3 months. Having an AI play a hundred games in the first three days of the leaderboard is a massive advantage compared to humans who are not awake 24/7, and would have only played a few dozen games (because they have to sleep and have lives).
This is why if you pull up the 1v1 leaderboards today, this model has already been passed by much better players who have a better winrate (71.9% vs 67.3%) with a similar number of games played (114 vs 107).
1
195
This is cool! Very interesting thing to note is that for specialized tasks, even for games like catan, beating a human does not actually require a huge model.
One thing I'm curious about was the env, I trained models for simpler but similar games in other competitions and the reward shaping with pure self play was super super hard, I had to write an algorithmic bot, then distill that bot to a model and use that as a warm start for rl. Would love to read a report w/ these details haha
We're excited to unveil VINCO, our frontier AI model trained to play and win 1v1 Catan.
We've secretly deployed it on the Colonist.io global leaderboard, where it has reached both the #1 and #2 positions, consistently beating every human world champion.
1
3
146
This feels to me like clickbait + mind boggling extrapolations
I feels quite naive to me to assume that 1.17b experiments will scale to quadrillion params? Chinchilla itself was like 70b?
&
if we have finite data then it's obvious as we scale parameter count the param count/token count ratio will rise anyways
Not to say that the research is illegitimate or bad/irreproducible though, just that it doesn't seem practical
Humans are ~1,000,000x more sample efficient than LLMs, and there are no good explanations.
We’ve found that a big part of the answer is that LLMs are *massively* underparameterized by many orders of magnitude. The brain has 100T synapses learning from ~200M tokens of language. Chinchilla prescribes a 1T-parameter model on 20T tokens. That’s 7 OOMs apart in parameters per token.
After a lot of painstaking work, we present new data-constrained scaling laws and learning algorithms to scale to (extrapolated) quadrillion-parameter transformers.
w/ @akshayvegesna, @EthanBaron78553, @asatheesh6561
Full paper led by @EthanBaron78553, with @ShmuelBerman, @andrewgwils
4
216
cool! i love the 5h limit removal. it helps with account planning a ton.
We’re renaming our Max plans:
Max 5x → Max
Max 20x → Ultra
Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do.
We’re also removing the 5-hour limit for verified Ultra accounts, starting today with most existing subscribers. Newer accounts keep a temporary limit until they’re verified, which helps us prevent fraud and abuse. You'll get an email when it's removed.
Weekly limits remain unchanged. View our plans here: claude.com/pricing
1
274
Let us cook :D
These results look quite legit to me
We've evaluated Gemini 4 Carbon and Gemini 4 Argon on FrontierCode 2.0, our coding benchmark that measure's a model's ability to go beyond the spec.
Gemini 4 Carbon scores near SOTA with a score of 40% and a cost of $4.04 per task, behind Claude Opus 5.5, GPT-6 Astra, and Claude Sonnet 5.5.
Gemini 4 Argon scores 26.3% with a cost of $6.41 per task, ahead of Muse Wonder 1.0, but behind Claude Haiku 5.5 and Grok 4.7.
See the leaderboard with full scores: cognition.com/frontiercode
2
126
I just found the time to read this and holy fk he's insane
Just the first bug of the chain was a universal UXSS LOL
1. visit attacker page
2. nav to another page
3. XSS
this is a very interesting exploitation idea, there are probably more places in chromium like this.
orz
Just dropped a new blog post! 🔥
blog.orange.tw/posts/2026-10…
2
293
kek retweeted
why pay for GPU-hours if what you need is tokens? just buy token-hours
Dedicated hourly throughput is LIVE on Touchmark.
Reserved in advance with SLAs, priced below on-demand.
Don't need the throughput? Sell it back.
touchmark.ai
If this works for codex like they claim, epic! Makes me wonder how big their caches are to be storing all of that on disk lol
Day 16/
You can now cache for 12h in the API and for logged in users across Codex & Work we keep cache forever and you don’t have any penalty for coming back to an old conversation or resuming after a few hours and are now guaranteed to hit the cache.
2
457























