The Two Ways the First AI-Enabled Attack Would Most Likely Happen

How much does it cost to make a leading AI model almost never refuse a dangerous request? In an Anthropic experiment cited by a new Belfer Center paper: about $4,400. Read the paper: https://www.belfercenter.org/research-anal... The paper runs a pre-mortem. It imagines that the first major AI-enabled cyber, biological or chemical terror attack has already happened, then works backward to ask which choices made it possible. It follows a model through five stages (development, deployment, elicitation, safeguards and detection) and checks the analysis against real cases of AI misuse. Its conclusion: such an attack would most likely trace back to one of two failures, either a closed-weight model shipped without state-of-the-art safeguards or an open-weight model whose safeguards did not survive tampering. It ends with nine falsifiable predictions and recommendations for AI developers, model distribution platforms, policymakers, and researchers and funders. ABOUT THIS VIDEO This video presents excerpts from "A Pre-mortem on The First Major AI-Enabled Terrorism Event and The Choices That Could Enable It: Plus 9 Predictions" by Max Kamachee, Jordi Weinstock, Antonia Juelich, Jonas Raedler, Elise Racine and Stephen Casper, published by the Belfer Center for Science and International Affairs, Harvard Kennedy School, in October 2026. All rights to the original work belong to its authors and publisher. I do not own this material and claim no credit for it; the research, analysis and conclusions are entirely theirs. This is an independent adaptation, not affiliated with or endorsed by the authors, the Belfer Center or Harvard University. Narration quotes and paraphrases the paper; citations and footnotes appear on screen. Chapters 0:00 Intro 1:18 What counts as a major AI-enabled attack 3:17 Failure point 1: Development 5:42 Failure point 2: Deployment 7:30 Failure point 3: Elicitation 9:40 Failure point 4: Safeguards 11:44 Failure point 5: Detection 13:48 Lessons from real AI misuse 16:51 The nine predictions 20:21 Two illustrative scenarios 22:34 Recommendations 24:23 Recap Credits Narration: AI voice generated with ElevenLabs. Music: generated with ElevenLabs Music. Visuals made for this video; figures and tables are reproduced from the paper. Spotted an error? Leave a comment and I'll pin a correction. #AISafety #AIPolicy #AIRisk