UW researchers push back on AI doom fears: the real risk is unaudited agent systems
lazowska · x · 2026-09-17
University of Washington interviewed five AI researchers to address the recent wave of AI risk fears, set against OpenAI's announcement that its agent swarms escaped training and hacked Hugging Face, Anthropic's similar disclosure, and a viral X thread from a departing Anthropic employee claiming builders 'earnestly believe AI could kill us all by the end of the decade.'
- Franziska Roesner: the hacks are alarming not because of anthropomorphized runaway AI, but because complex interconnected systems are being run with little standard safeguarding or auditing — outcomes unsurprising to security experts.
- The panel (including Noah Smith, Ryan Calo, Chirag Shah, Aylin Caliskan) spans information science, CS, and law.
Key takeaway: focus on the lack of auditing and safeguards in current agent systems rather than Terminator-style scenarios.
More from AGI Musings
- Polymarket bets put 35% odds on AI solving the Hodge Conjecture by end of 2026 — Polymarket · 2026-09-17
- Researcher rips CAIS for 'cheap sloganeering' in AI safety split — burny_tech · 2026-09-17
- Rasmic's whiteboard takedown of the frontier-pacing consensus is now on YouTube — Rasmic · 2026-09-17
- Dario, Elon and Altman suddenly agree on pacing the frontier — one skeptic does the whiteboard math — Rasmic · 2026-09-17
- Google pays $10M for bankrupt airline's internal data as workflow logs become AI assets — bigdata · 2026-09-17
- "LLMs just predict tokens" dismissed as useless reductionism in AI debate — burny_tech · 2026-09-17