
breaking and attributing LLMs @ https://nitter.cf/t.co/MNK0quwU7G curr: student researcher @googledeepmind prev: ML intern @ambosstech
London, UK
Joined December 2020
- Tweets46
- Following332
- Followers74
- Likes208
Pinned Tweet
Jailbreaking research assumes a fixed model.
Test-time training breaks that assumption: the attacker gets a lever on the weights, not just the prompt.
We show this opens a new attack surface that undermines LLM safety guardrails.
🧵 (1/6)
Simone Antonelli retweeted
Today we are happy to share one of our first open research releases: Latent Sparse Attention (LSA), a step towards subquadratic attention that matches accuracy of full attention at ultra-long context.
Sparse attention attends to only a few tokens per query, but it has to find the right ones. At 10M tokens, that search takes 98% of the attention layer's time. LSA makes sparse attention up to 22.7× faster and full attention 72.4x faster.
With @kubicek_ales @caciolai @PengLiangzu @SukritSumant @Ben_Hoov @FraserGreenlee @RMBattleday @radim_urban @thoefler @jcrwhittington
Simone Antonelli retweeted
🐬 OpenWhistle is Spotlight @ NeurIPS2026 (top 1.3%)!
We release the largest open dataset of dolphin whistles (114h), enabling ML at scale to better understand one of the most complex non-human communication systems.
Led by @FMustun & @chiara_semenzin
📄 arxiv.org/abs/2609.34839
Simone Antonelli retweeted
One photo in, a complete 3D object out, faithful in shape, color & pose.
GenIA grounds a frozen 3D foundation model (SAM3D) in the input images at test time. Works from single images, multi-view, even monocular video 🧵
🌐 facebookresearch.github.io/G…
Simone Antonelli retweeted
We were very fortunate to host an incredible group of people at our workshop on agentic privacy and security at Google and are now eager to show the results of our discussions in this report 🧵
Simone Antonelli retweeted
Small bird, fast wings, Kolibri is here.
78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.
Now the weights are yours. Run it on your own hardware, under Apache 2.0.
Simone Antonelli retweeted
Introducing Limite 1B - Violetto.
A model for high-frequency mathematical intelligence.
Simone Antonelli retweeted
Meta just released OmnilingualGAIA2 on Hugging Face
A multilingual extension of the GAIA2 agentic benchmark -
6400 scenarios across 10 languages to test AI agents.
Simone Antonelli retweeted
Jailbreaking research assumes a fixed model.
Test-time training breaks that assumption: the attacker gets a lever on the weights, not just the prompt.
We show this opens a new attack surface that undermines LLM safety guardrails.
🧵 (1/6)
Simone Antonelli retweeted
Can a sentence carry a sound?
In Communicating Sound Through Natural Language, we introduce lexical acoustic coding (LAC): a way for LLM agents to transmit short sounds as structured English, then re-render the same audio back from that text.
(1/6)
Our first defense flags TTT via a perplexity shift on a private holdout. It catches our attacks, but robust safety needs dynamic alignment.
Joint work with @MS_Akhondzadeh and @abojchevski
📄 arxiv.org/abs/2605.22984
💻 github.com/UoC-tail/ttt-jail…
🌐 uoc-tail.github.io/ttt-jailb…
(6/6)
1
2
174
On a personal note: I'm currently in London for an internship at @GoogleDeepMind and will be spending the summer here. Hit me up to chat about AI safety, from LLMs to agents.
4
83
Simone Antonelli retweeted
🎨Activation steering can reliably push a text-to-image generator toward a visual concept, but at a cost: each concept needs its own estimation.
⚡HyperTransport (HT) predicts the intervention directly, matching per-concept SOTA at 3–4 orders of magnitude less cost.
[1/6]
Simone Antonelli retweeted
🎉 7 new ELLIS Units join our network
Units Franconia 🇩🇪, Grenoble 🇫🇷, and NRW 🇩🇪 strengthen existing hubs; Units Slovenia 🇸🇮and Sweden 🇸🇪 add new countries; Units Czechia 🇨🇿& Denmark 🇩🇰 expand from city to national Units.
Learn more: bit.ly/4nN3vcD
Simone Antonelli retweeted
Force closures in the Lightning Network reduce reliability and capital efficiency. Check out our latest research on predicting force closures in the Lightning Network using ML!
Simone Antonelli retweeted
The Lightning Network makes Bitcoin payments faster and cheaper, but forced closures remain a major challenge: they reduce reliability and lock up capital.
Check out our new work with @sntonelli and @ambosstech on predicting forced closures in the Lightning Network ⚡️
















