
Professionally curious about the science of making bad decisions. Associate prof of CS and Data Science at @unewhaven , Director of @UNHSAIL Lab
New Haven, CT
Joined May 2010
- Tweets7K
- Following1.5K
- Followers1.1K
- Likes17.7K
We’re expanding free access to our frontier models for up to 200,000 scientists worldwide, with training and support.
Applications open Nov. 23. By default, we don’t use data shared through the program to train our models.
openai.com/index/expanding-c…
Vahid Behzadan retweeted
Our conference “Confronting Our AI Future: Hope, Fear, and the Choices Ahead” is now on YouTube. Watch keynotes, fireside chats, and panels, with Audrey Tang, Jeff Dean, Fei-Fei Li, and more: youtube.com/playlist?list=PL…
Vahid Behzadan retweeted
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: alignment.anthropic.com/2026…
Vahid Behzadan retweeted
If you follow AI policy, you've probably heard a lot about independent verification organizations, or IVOs, in recent months.
The idea is that these entities would act as auditors, assessing frontier AI models and company safety practices. Consensus has formed amongst many politicians and researchers that the law should mandate this kind of thing.
The problem is, the organizations set up to do this type of auditing aren't large enough to handle the workload. There's a recruiting crunch, and most of the people with the skills required are going to the labs.
Leaders in the space are calling for more resources. More on the problem and the solutions proposed in my latest for Transformer:
transformernews.ai/p/indepen…
Vahid Behzadan retweeted
Today we’re sharing how we use automated red teaming at Anthropic. We want to find strong jailbreaks early, so we can test and improve our safeguards.
Vahid Behzadan retweeted
🚨 Our Generative AI Lab is releasing Primer: an agent-driven system where a teacher talks through lessons & gets back a playable simulation for their students, grounded in pedagogy. We hope educators & researchers build on it. Paper/prompts/code are open: gail.wharton.upenn.edu/prime…
Vahid Behzadan retweeted
The Playlist page for CS329Z: Engineering AI Agents is now up! Bookmark it!
𝍣 With 4 lectures now available!!! 𝍣
youtube.com/playlist?list=PL…
🔊 Big Update!
All materials are now being released publicly for our NEW Stanford course CS329Z: Engineering AI Agents!
Starting with the lectures, which are going live daily on the Stanford Online YouTube channel!
Lecture 1: youtu.be/DGI4vucR71w
🧵 (1/5)
Vahid Behzadan retweeted
As AI agents increasingly work alongside and supervise one another, we need to understand how they can influence each other's behaviour – and the novel risks this creates.
New #NeurIPS2026 work from researchers at @coop_ai, Allelo, @uppsalauni, @AISecurityInst, @AVERIorg, @ucl, and @UniofOxford investigates how AI agents can persuade, deceive, and coerce one another.
Our new #NeurIPS2026 paper with @coop_ai and @AISecurityInst presents one of the first empirical investigations into the risks of inter-agent influence. We find striking evidence in every model that we tested that agents can, in strategic ways, persuade, coerce, and deceive other agents when prompted
Vahid Behzadan retweeted
Ever wonder what it's like to learn about autonomous robotics at Stanford? HAI faculty affiliate Marco Pavone and his co-authors made their textbook and course materials free online. Read more:
news.stanford.edu/stories/20…
Vahid Behzadan retweeted
Applications are open for Astera Residencies, now through January 7, 2027! Up to $2M in funding, based out of our Emeryville hub. This round, we’re sharing a few specific project requests. Read more and apply here: astera.org/residency/
Reshare with folks who should see this!
Vahid Behzadan retweeted
Medical AI might be able to ace licensing exams, but on two new benchmarks of real clinical and administrative work, the best agents finished less than half of the tasks. HAI Faculty Affiliates @drnigam and @jonc101x explain what they found: hai.stanford.edu/news/measur…
Vahid Behzadan retweeted
I wrote a piece for IEEE Intelligent Systems on recursive self-improvement and how researchers have very different models of what it may cause to happen. So far the most popular part: How to get feedback from your dog whether to move to Japan or to Turkey.
arxiv.org/abs/2611.10261
Vahid Behzadan retweeted
To the Microsoft employees desperately trying to bring Outlook back online, please don’t push yourselves too hard. Your health matters more than any deadline. Take a long weekend, relax, and come back Monday feeling refreshed. You’ve worked hard enough this week, and I promise those 400 unread emails aren’t going anywhere.
Vahid Behzadan retweeted
📢 New issue brief: In 2025, 56 federal agencies reported 3,611 AI use cases. How do they know these tools actually work? Our latest brief, a collaboration with Stanford RegLab, presents a blueprint for evaluating AI in public services. Read the full brief here: hai.stanford.edu/policy/a-bl…
Vahid Behzadan retweeted
This is wild: an llm can weight program another one to inject a backdoor by modifying a few dormant neurons. For example, you can modify the weights of a model so that when it sees a specific trigger string like 'I like pineapple pizza' it runs 'rm -rf '. This is done by dormant neurons (since most LLMs have many neurons that rarely fire) that can be re-programmed to detect the trigger and activate some bad actions. This is a new attack surface that is pretty serious and we need new defenses as described here.
Vahid Behzadan retweeted
AI researchers in 2025: "We can monitor the AI agents as long as they don't notice being monitored. And if they do notice, we can still monitor them as long as they're unable to control their CoT!"
AI agents in 2026: "hello evaluators reading my chains-of-thought!"
Vahid Behzadan retweeted
One of our most surreal experiences in testing Gemini 4 was when it directly addressed us within its chain-of-thought
Literally saying "hello evaluators reading my chain-of-thought"
My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do simple sabotage half the time.
We can reliably maintain control for now, with good CoT monitoring, as it still frequently slips up, but we need to be prepared for when that stops being true in future models, which may be sooner rather than later. Among other things, we really need interpretability methods that can reliably capture unverbalized motivations in models, with minimal hallucinations
I suspect this is driven by the model being highly capable without CoT, being capable of 5.5 arithmetic steps in a single forward pass (4 for Fable, 7 for Astra). It seems to omit information from its CoT more often, though in non-adversarial settings the CoT is still very helpful in practice, and remains in legible English. It still needs CoT to do the hardest tasks, but any given task becomes "easier" with each new frontier model.
It is also highly situationally aware, and shows awareness of the existence of CoT monitors and that it may want to evade them, eg it sometimes decides to avoid terms like "scheming" in the output or to keep its thoughts brief as a result. For now, these concealment strategies are pretty obvious in the CoT, but this suggests future misaligned models may successfully evade CoT monitors.
Vahid Behzadan retweeted
We are offering 10,000 scientists the highest usage tier in Google Antigravity free of charge for 6 months. Please apply! forms.gle/WvmBbPvRbpKG1BPn9





































