Excited to announce the 2nd edition of the Cryptography and Machine Learning Winter School, taking place in Milan on January 25–28, 2027!
Registration is now open and free of charge! Hope to see you there :)
bidsa.unibocconi.eu/cryptogr…
Excited to announce the 2nd edition of the Cryptography and Machine Learning Winter School, taking place in Milan on January 25–28, 2027!
Registration is now open and free of charge! Hope to see you there :)
bidsa.unibocconi.eu/cryptogr…
I matriculated at Oxford this week in a part-time, remote DPhil in Engineering Science with @casdewitt, @RobertTrager and @maosbot!
I’ll be working on frontier AI auditing and multi-agent risk, while continuing full-time as Director of Standards at @AVERIorg. Looking forward to what's ahead.
Really proud of this work, and of the leads @BenHagag20 and @wlanderson0.
ORBIT started from a simple question: how do we systematically study security once the object of evaluation is a multi-agent system, rather than a single model?
Lots more to do, but very excited to have this out.
New @NeurIPSConf paper: we're releasing ORBIT: A Framework for Multi-Agent Safety and Security Evaluations. Open source, with configurable attacks & defenses and dynamic topologies, all built on @AISecurityInst's Inspect.
When we started this work a year ago, we didn't imagine how urgent it would become by the time we published. Below are some of the interesting results and where we think this should go next. If you want to talk it through, let us know 🧵
Fantastic, systematic work on inter-agent influence just accepted to NeurIPS'26 - congrats to lead @ChandlerDSmith , @prpaskov as well as senior author @lrhammond and the others
Right now we learn how AI agents behave together after the fact, through hacks, logs, and headlines.
Soon agents will interact with other agents on our behalf. Commerce will rely on those exchanges. How easily can agents persuade, deceive, and coerce each other?
Better to find out in evaluations than in the logs. We started measuring.
NeurIPS 2026, paper coming soon
With @ChandlerDSmith Qi Guo @tilli_cecilia@SophiaHatz4@DavidDAfrica@lrhammond@casdewitt@philiptorr
Right now we learn how AI agents behave together after the fact, through hacks, logs, and headlines.
Soon agents will interact with other agents on our behalf. Commerce will rely on those exchanges. How easily can agents persuade, deceive, and coerce each other?
Better to find out in evaluations than in the logs. We started measuring.
NeurIPS 2026, paper coming soon
With @ChandlerDSmith Qi Guo @tilli_cecilia@SophiaHatz4@DavidDAfrica@lrhammond@casdewitt@philiptorr
METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.
Super excited about co-hosting the AI4GOOD workshop at @NeurIPSConf in Paris - there will be a dedicated multi-agent security/safety track as well in collaboration with @krawiecka_kl@swapneel_mehta and others - the submission deadline is approaching fast!
We are so excited to host the 2nd edition of the Trustworthy AI for Good (AI4GOOD) Workshop at #NeurIPS2026 in Paris, France 🇫🇷
We aim to bring together the AI safety, AI for social good, and AI policy/governance communities to connect what models can do as individual systems with what they do when deployed across populations.
For full details, welcome to visit our website at:
🔗 trustworthy-ai-for-good.gith…
📍December 12, 2026 · Paris 🇫🇷
We have an amazing lineup of speakers and plans. More details 🧵
Our work from 2024 (started by Andis Draguns and @dyushag) shows that we can construct LLM backdoors where finding the trigger is cryptographically hard. Next month, we'll release a paper where a full set of weights (and therefore, capabilities) can be hidden as a backdoor that's cryptographically hard to unlock. This could have some consequences for AI security...
yes, although there are backdoors where finding the trigger requires reversing a SHA-256 hash (caveat: this method is not robust to noise/fine-tuning, and it might still be possible to detect that the model *has* a backdoor)
arxiv.org/abs/2406.02619
Proud to help launch Orbit v0 — a framework for multi-agent evaluations based on @AISecurityInst's Inspect, starting with security but built for safety and capability research too. Huge team effort, and excited to see where this goes. Hugecongrats to @BenHagag20@wlanderson0 who have been developing this as part of my @MATSprogram, and @csrijac for mentoring. Stoked @coop_ai will be stewarding this going forward.
Today we're releasing Orbit v0: a framework for multi-agent safety and security research.
If you've tried running multi-agent experiments, you know the drill - weeks of infra before your first result. Built on Inspect, Orbit lets you swap environments, threat models, and topologies without rebuilding anything.
Work with @wlanderson0, @casdewitt, Sarah Scheffler as part of @matsprogram. We’re presenting ‘Architecture Matters for Multi-Agent Security’ at @icmlconf this week in Seoul, come find us!
If models can think for 100,000 tokens, why do they still lose the plot?
Come join us for this AI4Science on alphaXiv talk: Long-Horizon Reasoning in LLMs.
In this session, Sumeet Motwani (@sumeetrm) and Charles London (@CharlieLondon02) will share recent work on both training and evaluating models that can reason over much longer chains of thought.
Their LongCoT benchmark tests whether models can handle long chains of dependent reasoning across different fields. Each step is solvable on its own, but the full problem requires planning, state tracking, backtracking, and avoiding compounding errors. Even the best models still score below 10%.
They will also discuss h1, which trains long-horizon reasoning by chaining short problems into longer dependency graphs, then using RL with outcome-only rewards and a gradually harder curriculum.
So if longer context windows are not enough, what does it actually take to make models reason reliably over long scientific and technical workflows?
Whether you’re working on frontier LLMs, AI4Science, reasoning, or just curious about what current models still cannot do, you should definitely check this talk out!
🗓 Friday May 15th 2026 · 11 AM PT
🎙 Featuring Sumeet Motwani and Charles London
💬 Casual Talk + Open Discussion
New mini experiment + blogpost + trajectories!
tldr; we boost performance of RLM(GPT-5.2) to double the best performing number (38.7% --> 65.6%) on LongCoT-mini without any training! An example of the mismanaged geniuses hypothesis (MGH) we (@zli11010, @lateinteraction) proposed earlier this month.
The LongCoT benchmark showed that frontier LMs and RLMs struggled to solve difficult compositional reasoning tasks. The paper generally attributes this to the RLMs inability to perform task decomposition, but we argue this is more our fault in how we prompt them; this capability is fully available to GPT-5.2 with an RLM harness!
Building on @raw_works's insightful blogpost and @sumeetrm / @CharlieLondon02 et al.'s incredibly useful benchmark, where they originally found RLMs to be incapable of solving the MATH and CS splits altogether. We did not train anything since the release of the initial benchmark.
To be fully transparent, these results are not meant to be added to their leaderboard either; benchmarks measure isolated capabilities, and we focus on showing (through different, rather specific prompting) that the capabilities required to solve these tasks are available to the models without additional training! It also has implications about how we would go about training these systems. Full blog below, it's a nice read :)