🚨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 .
Post
Conversation
Post your reply
The threat isn't hypothetical. There have already been a number of violent and deadly incidents involving small-scale terror and criminal misuse of AI. Frontier AI companies and world governments treat the threat of a bio/chem/cyber attack enabled by AI as a credible concern.
From an AI safety science perspective, this paper considers how (1) these harmful capabilities would be developed, (2) they would be deployed, (3) the dangerous capability would be elicited, (4) safeguards would fail, and (5) detection would fail.
We also look at how harmful capabilities have proliferated and been misused in the past several years, involving (1) non-consensual AI nudification, (2) AI-assisted cybercrime, and (3) AI-uplifted small-scale terrorism events.
This leads us to our central prediction: Such an event will trace to either a closed-weight model deployed without state-of-the-art safeguards or an open-weight model.
We make 8 other predictions as well: three about the closed-weight model pathway to harm, three about the open-weight model pathway to harm, and two about either pathway.
Finally, we describe two grounded, realistic illustrative scenarios for how such an attack might occur, involving:
1. A non-state group, a closed-weight model, and a biological attack
2. A state-sponsored group, an open-weight model, and a cyberattack
We hope that this paper can help ground serious discussions about mitigating AI-enabled terrorism risk.
Paper (coming to arXiv/SSRN soon):
And a piece by the Harvard Gazette:
Props to for the extraordinary talent, persistence, and vision in leading this paper.
And a big thanks to other coauthors: Jordi Weinstock, , Jonas Raedler, & Elise Racine.
Work done at and at .
Runway is now in GPT-6 Astra. Ask for a video in ChatGPT and Runway makes it. Multi-shot, same cast, in the thread.
has been sounding the alarm on this as well.
Quote
Chief Intelligence Officer
@SINewsBriefing
Robert Pape (@ProfessorPape) says it hasn't happened yet, but he's concerned that ChatGPT, Claude or DeepSeek will empower a lone wolf assassin via “extremely good tactical programs that they can execute. And they're going to work through the defenses.” x.com/MarioNawfal/st
The media could not be played.
I am now able to make an equally bold prediction:
There will be a cyber attack in the next several years utilizing sophisticated “personal computers” and “fiber optic networking” that can currently be purchased with no restrictions (!!!!!)
belfercenter.org/sites/default/ Now have skimmed. Good going, high marks for state of art including appendix D
Thanks so much. Fwiw, a new docufilm👇gets warnings about unchecked AI right, even if it misdiagnoses the impulses driving AI's breathtaking pace and direction.
you nowhere establish or justify why 2027 / 28 is the relevant window. more like "we think...so there might be".
and most of predictions are literally one prediction "models will be models"
which of the 9 predictions would you bet resolves first? curious too whether the 2 scenarios run through open weights or a jailbroken api model
Which of the nine predictions would give the most useful early warning, while there's still time to intervene?
Super important paper. Thank you for sharing. Needs to be widely read and appreciated.
would love to chat about this work! any chance you'd be up for a call?