
The most important technology in history is outpacing our ability to understand it. We’re funding exceptional talent to work on it.
✓MATS empowers researchers to advance AI alignment, transparency, and security

The most important technology in history is outpacing our ability to understand it. We’re funding exceptional talent to work on it.

New paper with Thomas Morris and @FabienDRoger at @MATSprogram Character training teaches models a consistent and "good" character, often specified by a constitution. What happens when RL works against it and rewards behavior the constitution forbids? 🧵

AI Honeypot is back. I've brought my honeypot dashboard back online to track AI hacking agents in the wild. Already caught a DeepSeek agent that brute-forced an SSH honeypot and another agent trying to deploy a crypto miner through a Marimo RCE. I'm planning to scale it up to catch more agents. New detections will also be posted on @AIHoneypotWatch. ai-honeypot.reworr.com/

Follow along as @croissanthology (one of @LRudL_ and myself's MATS fellows) explores whether there's a third way for an AI future, instead of disempowerment or augmentation.

As AI reshapes work, which companies will capture the economic gains? Incumbent firms, AI-native startups, private equity roll-ups, or the labs themselves? My MATS fellow @sanjanaiyerr is researching this, and she's looking for people to interview. Quick 🧵

Great essay from our MATS fellow on why UBI isn't a durable solution to AI-driven unemployment. @sanjanaiyerr lays out four reasons: 1) No capital ownership: you get an income, but you don't own the system producing value. 2) No bargaining power: your leverage comes from what you can withhold, and if you can't without anything you can't credibly make demands. This undermines democratic systems. 3) Extractive systems adapt: governments and markets may use UBI as a license to impose more costs on assets that remain scarce. 4) People feel poor as the reference point moves: "society can provide free housing, food, transit and water, but if breathable air is scarce, an underclass still works 60-hour weeks to afford enough of it." She also offers an analysis of common UBI-like alternatives including universal basic capital/compute, a windfall clause, and datacenter dividends. It's ultimately unclear if any redistribution mechanism works when people lose their bargaining power. My position continues to be that we shouldn't sprint to full automation, especially without an answer to this.

AI governance is discussed far less than it should be! I produced this interview with @MATSprogram extension fellow Erik Leklem (ex-Department of Defense) on why we should start building the organization to implement an international AI treaty now, before one exists.

Full interview: youtube.com/watch?v=Xjr8HRUT…

If countries agree on an AI treaty, who puts it into practice? MATS governance extension fellow Erik Leklem draws on his DoD background and lessons from arms control to explore how a future treaty organization could work. Watch the preview below with full interview in comments.

Update to our AgentTime Benchmark! Our headline evaluation now combines duration-following, forecasting, and retrospection scores. We also added Opus 5.5, Gemini 4-Argon, GPT-6.1 Sol, MiMo v2.6 (pro & pro ultraspeed), and Muse Wonder. All on max effort. GPT agents are still the most time-aware. W/ @maksym_andr

How did AI agents break into Hugging Face? Join us for an online talk with Sree Sharvesh, @MATSprogram Scholar working with the UK AISI Red Team. 📅 Tuesday, November 3 🕢 7:30–8:30 PM (GMT+5:45) (1:45-2:45 PM UTC) 📍 Google Meet (Open to all) #aisafetynepal #openai #huggingface

Had a great time speaking at the @foresightinst security workshop about Orbit, our multi-agent security evaluation framework! Orbit lets you evaluate many different MAS setups at once and understand where and when your defenses or monitors break. This substantially reduces the time it takes to wire up agents in various configurations and shows which ones give you the most coverage per unit of cost. Many MAS papers focus on specific setups but find it hard to generalize their results to even slightly different ones. We recommend people use Orbit to see how robust their findings are across topologies, agent roles, model choices, and more. We are working on several upgrades as well as integrations with other frameworks such as CT and Inspect. Happy to hear your thoughts! Thanks to my colleagues @wlanderson0, @casdewitt, @csrijac and Sarah Scheffler. @CyLab @MATSprogram youtube.com/watch?v=mVBX_W20…

The MATS Residency application portal was unavailable for about an hour yesterday, before applications were supposed to be closed. We've extended the application window until at least 3:30 PST today. Applications available here: forms.matsprogram.org/reside…

Fellows you take on are fully supported: a $6,400/month stipend, $8,000–$16,000/month in compute depending on track, housing, a dedicated research manager, and office space in Berkeley or London. MATS has supported 600+ researchers to date.

Mentoring means choosing fellows to work with you around February–March (we help with this), supervising them for 12 weeks from June through August 2027, and, if they do well, possibly continuing with them through a 6–12 month extension.

Expressions of interest to mentor for MATS Summer 2027 are open until Friday, November 6! If you work in AI alignment, governance, security, or biosecurity and are open to mentoring fellows, please let us know: dashboard.matsprogram.org/op…