Anto

We won 1st!! @usmannk @juancito @DrasticWM & adriro . Thank you @wonderland for organizing. We had so much fun

@ar0cket1 Thanks for sharing :) will run some experiments

@blocksec Well RL scaling works with a basically unlimited upper bound until your saturate your capacity. Scaling the base helps the RL scaling do better. The new Mythos checkpoint is more RL scaling for instance

imo this big jump is not really a research jump from the current line but just scaling. I think that we would very much be on the line (or not as far off), if GPT 5.5 and Mythos were the same size as their previous models (5.4 and Opus)

@ar0cket1 I wonder if the same base model with RL can hillclimb or does it need a larger base too :)

@blocksec More of a capacity thing than data, got 5.5 and mythos are larger than the previous trend and thus come with stronger than previous trend improvements. You can’t tts a weaker model like Kimi to do nearly as well as a stronger model. The tts is bounded by raw capabilities

@ar0cket1 Okay , so a larger pre-train. When backtesting on larger codebases , i did see improvements in performance with tts, my intuition was that the pretrain had enough data and tts helps access that ?

@blocksec This isn’t what I’m saying, I’m saying that the shift off the trend line are rather from just model param size scaling that has happened recently. Also Kimi likely wouldn’t come close to this. You can only test time scale a little until you saturate it

Opus this, GPT that bro, it all depends on your task at hand and whether it is in the model's distribution. If it is, it's great; if it isn't, it's just okay. There is no overall better model; it depends on your task.

@hrkrshnn This framing makes sense, but i would imagine there is a lot more no Claude code ppl out there? I would think that’s the 90%

OpenClaw isn't hype. There was an entirely different audience that wasn't on Claude Code or Codex for whom OpenClaw gave the right interface. Both Anthropic and OpenAI shipped products that were inspired by openclaw.

An email is a single logprob update. A conversation is online learning with back-prop.

@simonw @_catwu Yes

@_catwu Is the model behind that regular Opus 4.7?

Claude Security is now in public beta, built into Claude Code on the web. Point it at a repo, get validated vulnerability findings, and fix them in the same place you're already writing code claude.com/product/claude…

Cool idea, but there are a few ways to get past it: root access on your mobile device or just point the camera at a screen 📸

To stop reward hacking on increasingly complex tasks, frontier labs will build secure systems.

@andreamichi This is awesome!

I’m excited about our results and very proud of the challenges our small (but mighty) team solved. We’ve already tackled a restricted context window that required us to learn summarization during training, and learned to show generalization from smart contracts to all types of vulnerabilities. depthfirst.com/post/dfs-mini1…

This week @depthfirstlabs introduced dfs-mini1, a security model trained via Reinforcement Learning to detect vulnerabilities in smart contracts. The model achieves pareto optimality on OpenAI’s EVMBench Detect and SOTA at pass@8 beating frontier models at a fraction of the cost

@Montyly @0xteddav AGI 🚀

Today I am releasing IsItVulnerable: a new tool I’ve been working on for the past several months: github.com/montyly/isItVu… It builds on recent LLM progress and over a decade of experience building security tools. I developed a new technique that combines abstract interpretation with machine learning The key insight is that this method abstracts the intelligence away entirely. I call it Abstract Intelligence, or AI The result is a major breakthrough in program analysis: IsItVulnerable finds all bugs with 100% recall Yes, all bugs. Fully guaranteed I have tested it extensively, and it has never failed. The results are honestly incredible April 1, 2026 marks a turning point for security, and the industry will never be the same My DMs are open for investors. Entry ticket starts at $500k.

@eigencloud 👀

if you build agents, tomorrow is a good day to be online

Iykyk

Music, food and security engineering. What do they all have in common? 🇮🇳 @blocksec from @eigencloud, tells us more 👇

I built this RL env for vulnerability detection tasks on smartcontracts using @hud_evals last night: github.com/antojoseph/sol…. You can improve open source model performance with GRPO and measure with evals!

All the big labs have their ai security products at this point. @OpenAI has advark aka codex security. @AnthropicAI just announced Claude code security. The next frontier is RFT. More on that soon!

New benchmark to test your web3 security agents against just dropped! Thanks to @paradigm and @OpenAI . You can find the bench here github.com/openai/frontie…

@hrkrshnn @OpenAI

Thanks @OpenAI!

Frontier models will become just a model in 3/6 months. Cybersecurity defenders have to use the frontier or you risk a hack you could have avoided! Cybersecurity for critical industries has changed for ever!