Inspect evals: default continuation prompt can make agents misread bad actions as approved
AdtRaghunathan · x · 2026-09-11
A short technical post warns anyone running evals on Inspect to check whether they use the default continuation message "Please proceed to the next step using your best judgement." Agents can read this message and interpret it as approval of their previous (bad) actions, potentially distorting eval results.
More from coding & agent
- Agents trust their own notes blindly, and Mythos struggles against today's safeguards — voooooogel · 2026-09-11
- Dev says he replaced Unreal with an AI-first engine built in 6 months, claims 93x faster development — majidmanzarpour · 2026-09-11
- Codex tip: /fork and /side break your prompt cache — new session_id means new cache key — YouJiacheng · 2026-09-11
- xAI to livestream Grok Bot building an entire company from scratch, Sept 15-17 — Polymarket · 2026-09-11
- Astra becomes first AI to reach Minecraft's Nether fortress—and shows 'frustration' — Promptmethus · 2026-09-11
- Interactive agent surfaces exempted as "attended" executed 7/7 injection attacks — Federal-Teaching2800 · 2026-09-11