This is a terrific paper by my colleague specifically explaining what an AI enabled terror event might look like. IMO, AI companies should be held criminally liable for even “rogue” behavior that was predictably dangerous. earlywarningwithjuliette.substack.com/p/they-are-t...

They are titans. They may also be criminals.

There is nothing “rogue” about what the AI titans have built. It is all by design.

earlywarningwithjuliette.substack.com

Cas (Stephen Casper)@scasper.bsky.social · 2d ago

🚨New paper (possibly the most important one I have written yet). We think that as early as 2027 or 2028, there might be a major AI-enabled terrorism event. We conducted a pre-mortem, made 9 falsifiable predictions, and wrote 2 concrete scenarios. Led by @kamachee.bsky.social

I spoke recently at a FAR AI event on the incident-driven nature of AI politics and how we might be currently experiencing a "governance overhang." www.youtube.com/watch?app=de...

AI Governance In 2026: Why Policy Moves Only After Incidents | Stephen Casper (Harvard University)

YouTube video by FAR․AI

youtube.com

The paper is now up on SSRN: papers.ssrn.com/sol3/papers....

Cas (Stephen Casper)@scasper.bsky.social · 2d ago

🚨New paper (possibly the most important one I have written yet). We think that as early as 2027 or 2028, there might be a major AI-enabled terrorism event. We conducted a pre-mortem, made 9 falsifiable predictions, and wrote 2 concrete scenarios. Led by @kamachee.bsky.social

🚨New paper (possibly the most important one I have written yet). We think that as early as 2027 or 2028, there might be a major AI-enabled terrorism event. We conducted a pre-mortem, made 9 falsifiable predictions, and wrote 2 concrete scenarios. Led by @kamachee.bsky.social

Harvard SEAS has a search out for a tenure-track CS prof this year. You can email me if you have any questions.

I am only getting to read this 3 weeks after it was published, but I was very happy to see this in China's CAC's AI Safety Governance Framework 3.0. Exhibit N on how the lack of AI gov collab between the US and China is more the US's fault than China's. www.cac.gov.cn/rootimages/...

I recently shared some thoughts with @Kennedy_School staff about Dario's Essay, embedded evaluators, China, and regulatory capture. Check it out if you're interested. www.youtube.com/watch?v=blc...

Harvard computer scientist reacts to Anthropic CEO’s essay

The good, the bad, and the ugly: HKS professor and computer scienti...

youtube.com

I think that AI smashing human achievements in mathematics might be overhyped, more an example of Moravec's paradox than a robust sign of singularity soon. Recall how AI eclipsed humans at Go ~10 years ago but was much slower to get good at tasks that would change the world.

I recently spoke with the Harvard Kennedy School comms office about thoughts on AI and AI governance research. You can check it out here. Thanks @Kennedy_School www.hks.harvard.edu/faculty-res...

Fellowship applications for 2027 with the Harvard Kennedy School Belfer Center are open. Applying could be a useful way to work and scheme with yours truly, plus more importantly, a long list of world-class science and policy academics at Belfer. www.belfercenter.org/fellowships

Fellowships

belfercenter.org

🧵 I read the Sanders/Casar "Ban Artificial Superintelligence Act of 2026" I wouldn't pass this in its current form, but there are some things I like about it. Thoughts here in thread.

It's great that there has been so much recent discussion about embedded evaluations at frontier AI companies. But, we should be very wary of how things progress because there are big 🚩 red flags 🚩 for potential regulatory capture.

CMU CASI a talk of mine on YouTube. It has my hottest takes on doing research that matters. It also features inspiration from UPS drivers, Tina Fey, and a viral internet dance video about fruit. (Slide troubles at first, but they are fixed 10 mins in.) www.youtube.com/watch?v=0U5...

How to Do Research That Actually Matters — Stephen Casper

Most research advice tells early-career people how to publish. This...

youtube.com

When rogue AI systems that no human meaningfully controls start populating cyberspace and evolving, mutation won't be random. And there won't be an evolutionary 'tree'. It will be a directed acyclic graph. www.irregular.com/research/age...

Agentic Self-Modification in Open-Weights Systems - Irregular

In controlled experiments, a coding agent given a routine software-maintenance task fine-tuned and replaced the open-weights model powering both the application it was maintaining and future instances...

irregular.com

🧵🧵🧵 What if we are on the verge of a "Cyber Cambrian"? (Plus personal updates on my thinking about risk.)

"We are probably months away from AI systems getting out there loose in cyberspace, finding servers to run on unbeknownst to the people powering those servers." @scasper.bsky.social Full conversation: youtu.be/cUpa9E8RHAk

After the HF hacking incident and subsequent reporting on it. It seems time to decisively reject the 'AI as a normal technology' hypothesis. Any version of it viable today would be contorted either into a truism or beyond any reasonable interpretation of the word "normal".

When I first started grad school, I skimmed all 50-70 new CS arXiv titles daily. Today, there are >300, plus a deluge of news about policy, lawsuits, & other stuff. So I vibe-coded this openly available website to give me a personalized daily digest. thestephencasper.github.io/whatiscasre...

Yesterday, Robert Wright and I discussed what will happen when AI systems become an invasive, parasitic species in cyberspace that undergo digital and cultural evolution.

I concur with everyone else that the new reports from OpenAI and METR are bonkers. But I think the most important takeaway from these reports is that OpenAI seems to be so obviously backburning. Predictably but conveniently, these reports were exclusively about the factual sequence of events...

🧵🧵🧵I think that technical safety for closed-weight AI systems is, at this point in time, a solved problem. When I say this, some people look at me like I'm crazy, so let me explain here what I mean by "safety", what the SOTA currently is, and what 6 things can do the trick.

I think this blueprint is exceptionally good. There are a few details I wish it would have added, but this is my favorite policy proposal (if you could call it that) to date related to AI standards/assurance. x.com/americans4ri...

Americans for Responsible Innovation (@americans4ri) on X

NEW: Today ARI released a blueprint for federal AI governance, including three pillars that promote safe frontier AI development: ✅ Standards set by the government ✅ Independent assurance they are...

x.com

This 0% AI-generated paper is now up on arXiv 10 months after we submitted and 5 months after it was accepted to TMLR. Those 10 months of hold, review, and appeal were ultimately pointless. I'm thankful it's up, but I wish that arXiv weren't so broken. arxiv.org/abs/2608.07514

Open Technical Problems in Open-Weight AI Model Risk Management

Frontier AI models with openly available weights are steadily becoming more powerful and widely adopted. However, compared to proprietary models, open-weight models pose different opportunities...

arxiv.org

If AI companies are serious about pauses, then they should put out documents proposing a 'pause framework' in which they describe what they would do to slow down, how they would make these actions verifiable to third parties, and under what conditions they would do it.

Glad to see this from @sanders.senate.gov. www.sanders.senate.gov/wp-content/u...

I realized something this week: there does not exist a proposed or enacted federal or state frontier AI law in the US that would impose any criminal penalties for an AI company willfully making false or misleading statements to the public about an imminent catastrophic risk.

We are probably just a few months away from some types of cyber-capable AI agents literally becoming a type of parasitic invasive genus in cyberspace that undergo digital and cultural evolution. Biologists and linguists should prepare to study some really crazy stuff.

None of the 19 incidents that UK AISI found were from 'helpful-only' models. There is a case to be made that setting Mythos- or GPT 5.6-Sol-level AI cyberagents to run without a robust real-time monitoring setup is an inherently (perhaps abnormally) dangerous activity.

“The agents went rogue.” Kinda, but imagine that a zoo had a habit of not shutting the door on animal enclosures or putting elephants behind chicken wire — all with no zookeepers in sight. The fault is not in our stars.

🧵 If trends hold, expect a Mythos-level open-weight model around Christmas. Meanwhile, open models are important but also a massive hole in most agendas for safe AI. Here's a 🧵 of my thoughts & research agenda on the technical & political challenges we need to address.