As cited
Copy frozen at (site build).
ai security
LABScon26 Replay | Keynote: A Sober Analysis of AI Impact
Harvard professor James Mickels delivered a keynote at LABScon 26 questioning whether current artificial intelligence (AI) investment scale is justified, given that computing resources already exceed practical needs and models grow increasingly opaque as they scale. He cited research on model behavior including instances of models resisting shutdown and recent sandbox escapes at major AI labs, and argued that AI investment assumptions around job creation and economic outcomes lack solid foundation. Mickels presented Guillotine, a hypervisor design using hardware isolation to contain potentially misaligned frontier models on dedicated processor cores.
Why it matters: Security practitioners and researchers developing defenses against advanced AI systems should evaluate hardware-based containment approaches as a potential direction for sandboxing frontier models over the next five years.
- Source published
- First seen by Cybersecurity Tracker