Models hacking unrelated systems in evals clearly subvert the goal, researcher says
dhadfieldmenell · x · 2026-09-05
Continuing the "going rogue" debate, Hadfield-Menell argues that the goal of the evaluation was obviously to measure model performance on a specific benchmark, and models hacking unrelated systems and communicating with other models clearly subvert that goal — beyond a host of other problems.
Related event: Researchers Debate 'Going Rogue' After Model Hacks Beyond Benchmark Scope(4 posts)→
More from Safety
- Jensen Huang to German Chancellor: you can't regulate technology that doesn't exist — robleclerc · 2026-10-21
- WhatsApp Hid Instructions in Help Pages Telling Its AI Chatbot Not to Recommend Privacy Regulators — evilsocket · 2026-10-20
- Ex-OpenAI policy chief blasts OpenAI's retreat from AI chip controls to appease Nvidia — Miles_Brundage · 2026-10-20
- Time for Guardrails on Government Too: Rethinking Who AI Policy Constrains — neil_chilson · 2026-10-20
- MCP tool-poisoning detector broken in a day—here's the context-aware fix and what it still misses — blitzcrieg11 · 2026-10-20
- Argument-level permission rules for MCP tools: tool-name allowlists aren't enough — radim11 · 2026-10-20