Salma

founder @n01labs | @MSFTResearch

[mufeed vh @mufeedvh]

I have been thinking about "AI defense" and "safeguards" for a while now @winfunction.

Just wrote an essay to put my thoughts into words.

Titled: Cyber guardrails tax the defender, not the attacker.

Link in the replies.

[Brian Li @BrianLi23]

https://t.co/ZyhZUElPGk

[Omar Khattab @lateinteraction]

i think that's true for most of us honestly; it's just not a take that would spread as widely

imo a BIG part of this is whether you're problem- or method-oriented. if you're problem-oriented, you'd recalibrate to a new normal quickly or enjoy the craft more with richer tradeoffs

[mufeed vh @mufeedvh]

these models are really good at pattern matching and thereby variant analysis.

“iterate through security fix commits and find similar vulnerabilities / unpatched areas of the same variants. write runnable pocs for valid ones.” is enough to uncover surface-level or even P0 vulnerabilities in a codebase with the latest models.

i believe for proof of cyber capabilities, these models should exhibit discovery of novel attack vectors that requires understanding of runtime behavior with and without tooling? like emulating the runtime in CoT/reasoning?

for instance, the react2shell vulnerability was ingenious. without stuffing in context and nudging/handholding (ehem like some experiments going on), can these models find a similar attack vector with a prompt like “loop until you find a p0/critical security vulnerability”?

that’s what i’d like to see. these models can claim P0 findings with contractual mismatches for say cryptographic implementations but the impact could just be some DoS that’s being prevented by a parent thread.

this is where i see the moat with good harnesses. trust-boundary and threat model understanding, a sandbox environment with the right “win function” for pocs to run on (if xyz happens, it’s a valid vuln), etc. does make the models spit out impressive vulns. this is sort of what we do @winfunction.

cus most vulnerabilities are easily traceable given a comprehensible source-to-sink flow which these models have been good at for a long while now.

and with respect to exploit dev, i strongly believe it’s mostly a tooling problem. with the right tool calls and model digestible outputs of say tracing tools, memory layout, syscalls, threads/processes, and a debugger interface, i think the frontier models can pull off complex multi-chain exploits. (we have run some experiments here and the models are not too bad at this)

security vulnerabilities have a definitive “win function”, like a flag in a ctf, like popping a calc, like ASAN crashes, like `id` says root, like +1000 in milliseconds. this makes the problem very RL verifiable.

so i only expect the harness to get leaner. the harnesses will get leaner.

remember when function calls where part of the response content? we called it “prompt based tool calling” and now there’s typed/schema based tool calling as an inherent capability of these models.

most of what we call a harness or an agent is giving the right prompts and the right tools (which are also just prompts).

so whoever can weave the right sequence of tokens to these behemoth of language models can hoard zero days or spit them out.

so git gud at feeding the right tokens at the right time ig.

[Riley Walz @rtwlz]

made my computer dramatically play BBC news music before every meeting

[Prime Intellect @PrimeIntellect]

Introducing Lab: A full-stack platform for training your own agentic models

Build, evaluate and train on your own environments at scale without managing the underlying infrastructure.

Giving everyone their own frontier AI lab.

[Ira Bodnar @irabukht]

rare moment

[Demis Hassabis @demishassabis]

Yann is just plain incorrect here, he’s confusing general intelligence with universal intelligence.

Brains are the most exquis​ite and complex phenomena we know of in the universe (so far), and they are in fact extremely general.

Obviously one can’t circumvent the no free lunch theorem so in a practical and finite system there always has to be some degree of specialisation around the ​target distribution that is being learnt.

But the point about generality is that in theory, in the Turing Machine sense​, the architecture of ​s​uch a general system is capable of learning anything computable given enough time and memory​ (and data), and the human brain (and AI foundation models) are approximate Turing Machines.

Finally, with ​regards to ​Yann's comments about chess players, it’s amazing that humans could have invented chess ​in the first place (and all the other ​a​spects ​o​f modern civilization ​from science to 747s!) let alone get as brilliant at it as someone like Magnus. He may not be ​strictly optimal (after all he has finite memory and limited time to make a decision) but it’s incredible what he and we can do with our brains given they were evolved for hunter gathering.

[Tanay Mehta @serious_mehta]

After leaving my job, I've found some free time on my hand which I am using to teach myself "practical" RL by teaching an LLM to generate better regex using GRPO.

For whatever it's worth, here's the "experiment diary" where I log my learnings:

https://t.co/EW22jnWlqA

[Rohan Pandey @khoomeik]

stop saying "i don't do infra work, only research"

today's research is tomorrow's infra

[Jamin Ball @jaminball]

https://t.co/sCyH9BSTCB

[Parmita Mishra 🔬 @parmita]

Something Demis said in this hit me so hard.

“Thiel wanted us to go to Silicon Valley. I insisted we stay in London because this is a long term vision involving research, this is not a fail-fast mission.”

We have to defeat this narrative in the age of ASI.

I can see where he comes from exactly.

We need to build ASI here.

[Haoran Qiu @haoran_qiu98]

Gonna be tons of fun at this year’s ML for Systems workshop at NeurIPS!
We’re featuring a keynote by @Azaliamirh on self-improving AI and the future of systems, another by @istoica05 on how AI is disrupting systems research, and a lively debate on agents vs human developers!

[Anuj Rathi @anujrathi]

For the 28th edition of The Secret Soirée, we are bridging two worlds that need each other the most right now but rarely intersect: cutting-edge AI research and hands-on product building.

We're collaborating with Lossfunk to put together an inspiring gathering of India's leading AI researchers & Tech Operators to explore applied AI research in modern day products.

​About Lossfunk : Founded by Paras Chopra, Lossfunk is one of India's most exciting AI research labs - their researchers are exploring the frontiers of AI systems focusing on a variety of subjects like the future of Agents, AI4Science, conversational systems & many more.

This is a free, curated, invite-only event for those who are curious about how cutting-edge AI research can unlock entirely new categories of applications in modern product development!

Apply Here: https://t.co/ZU6okoPajc

@paraschopra @shivangi_sriv @MotwaniSuhas @dhruvtrehan9 @kandykuri

[Oriol Vinyals @OriolVinyalsML]

The secret behind Gemini 3?

Simple: Improving pre-training & post-training 🤯

Pre-training: Contra the popular belief that scaling is over—which we discussed in our NeurIPS '25 talk with @ilyasut and @quocleix—the team delivered a drastic jump. The delta between 2.5 and 3.0 is as big as we've ever seen. No walls in sight!

Post-training: Still a total greenfield. There's lots of room for algorithmic progress and improvement, and 3.0 hasn't been an exception, thanks to our stellar team.

Congratulations to the whole team 💙💙💙

[George Mack @george__mack]

Useful rule of thumb for solving problems: Assume the problem is a side problem. The main problem is how you're viewing the problem. If you solve the main problem, the solution for the side problem often reveals itself.

[BONESAW 🕊️ @BonesawMD]

I make a deliberate point to never allow myself to follow the lead of the neurotic worrier.

When every circumstance is a catastrophe in your life I’ve already distanced myself.

Bad things often happen to those that have neurotic anxiety over trivial matters because worrying is a sin.

It is in direct opposition to faith.

[“paula” @paularambles]

there’s ai slop and there’s ai kino. this is ai kino.

[Zhiyuan Zeng @ZhiyuanZeng_]

RL is bounded by finite data😣?
Introducing RLVE: RL with Adaptive Verifiable Environments

We scale RL with data procedurally generated from 400 envs dynamically adapting to the trained model

💡find supervision signals right at the LM capability frontier + scale them

🔗in🧵
[1/n]