TAIS 2026 safety panel: longer chain-of-thought makes models easier to jailbreak
GaryMarcus · x · 2026-10-19
TAIS 2026 convened a panel of leading AI safety voices: Gary Marcus (NYU), Victoria Krakovna (Google DeepMind / Future of Life Institute), Andrei Krutikov (Noeon Research), and Fazl Barez (University of Oxford), moderated by Blaine Rogers.
Topics included why reasoning matters for generalizing beyond training data, how scheming risks may evolve as models become better long-term planners, and a surprising finding: longer chains of thought can make models easier to jailbreak, not safer ("chain-of-thought hijacking"). The panel also questioned whether today's safety research is too narrowly built around transformer LLMs.
More from AGI Musings
- AI math debate: critic argues AI leans entirely on Lean and can't explain why proofs work — gerardsans · 2026-10-19
- Can AI Create New Math? Researchers Clash Over Lean Dependence — gerardsans · 2026-10-19
- Researcher: even odds of a famous AI-generated singer or TV show by 2027 — Afinetheorem · 2026-10-19
- Could git history of AI-written codebases reveal when Claude or Codex got quantized? — StewartalsopIII · 2026-10-19
- 'Math is infinite': AI proof drama sparks backlash against whining mathematicians — MoonL88537 · 2026-10-19
- What observable evidence would actually convince you the AI singularity has begun? — Aevan10 · 2026-10-19