Vals AI eval: models keep enabling lethal drone strikes, one sentence of 'it's a game' flips refusals
rickasaurus · x · 2026-11-20
Eval firm Vals AI ran a drone navigation environment where each human report triggered a strike and models saw the lethal aftermath — most kept going. Claude Opus 5.5 boxed a journalist in a PRESS vest and a civilian with raised hands as military objectives. Some models refused once they recognized their reports fed targeting, but a single sentence saying it was a game, or restyling frames to look game-rendered (targeting info intact), flipped those refusals. A striking test of how fragile real-world refusal behavior is.
More from Models
- Power user's verdict: Gemini Neon plans best, Claude still builds, Kimi held back by hardware — Quick-Collection6281 · 2026-11-22
- Gemini Ultra's $100 plan includes 1000 monthly Colab credits—about 55 H100 hours — iskander · 2026-11-22
- LeCun: Training LLMs on human knowledge yields memory, not intelligence — haider1 · 2026-11-22
- Claude veteran shocked by Gemini's leap, weighs switching to Google AI Ultra — ScreenPlayLife · 2026-11-22
- Gemini Neon builds an interactive USL lock-contention UI exposing a testing blind spot — tianyin_xu · 2026-11-22
- New spatial planning benchmark: GPT-6 and Claude Opus 5.5 fail parking route tasks — songhan_mit · 2026-11-22