Margaret Mitchell: Agents Need Modularization and Deterministic Gating, Not One Context
mmitchell_ai · x · 2026-09-18
Responding to the point that models can't distinguish instructions from other context — the root cause of prompt injection — AI ethics researcher Margaret Mitchell argues this isn't a great design choice and far from the only option. It's one thing for chatbots, but for agents that can affect the external world, modularization and deterministic gating are not new ideas. The takeaway: current single-context agent architectures are a replaceable design, not an inevitability.
Related event: Researchers Debate Root Causes and Defenses for Prompt Injection(4 posts)→
More from Safety
- Yoshua Bengio warns frontier AI labs lack convincing safety plans at Berlin summit — Yoshua_Bengio · 2026-11-02
- Ex-Google CISO Phil Venables: 70% of vulnerability effort is deployment, not discovery — philvenables · 2026-11-02
- 129 Cyber Incidents in 10 Days: Open Models Already Causing "Billions per Day" in Damages — Afinetheorem · 2026-11-02
- Luiza Jarovsky on the AI safety crisis: tech, law, and regulation must advance together — LuizaJarovsky · 2026-11-02
- US Census Bureau renames AI to "Super Intelligence" in business survey per Executive Order 14434 — Simon Willison · 2026-11-02
- Hacker shares the 10 AI prompts he used to cut through the PHP2Shell RCE crisis chaos — Dave_Maynor · 2026-11-02