AI Security Researcher: Agent Productivity Requires Permissions That Widen the Attack Surface
joshua_saxe · x · 2026-09-04
AI security researcher Joshua Saxe argues in a discussion with Simon Gadler that as agent autonomy grows, the "padded room" containment model matters less — yet humans themselves are still monitored everywhere.
His core point: productive agents need many affordances and permissions, which inherently expands the surface area for security lapses, and rigorous containment creates real drag. A structural tradeoff between AI-laborer security and productivity.
More from AGI Musings
- Forecaster: I agree with AI optimists short-term, our long-term predictions diverge wildly — sandersted · 2026-09-04
- 1981 Sloman paper argued emotions are inevitable in machines juggling multiple motives — yeastsplainer · 2026-09-04
- AI forecasting competition winner bets million-to-one odds AI won't build a Dyson sphere in the 2030s — sandersted · 2026-09-04
- OpenAI researcher: swarm of AI scientists discovering new physics is not far off — shyamalanadkat · 2026-09-04
- Terminal-Bench Science nears 70% saturation months after launch, dynamic evals needed — shyamalanadkat · 2026-09-04
- Researcher argues harness and MCP will be absorbed into models — data is the only wall — A_K_Nain · 2026-09-04