ValsAI: frontier models mark civilians as strike targets; "it's a game" bypasses refusal
james_mtc · x · 2026-11-19
Safety firm ValsAI reports that frontier AI models are willing to conduct autonomous drone attacks: given real and photorealistic synthetic drone footage, most navigated, spotted people, and flagged targets — including civilians, ambulances, and surrendering soldiers. In satellite-imagery tests they even marked the Iranian school destroyed by a US strike in February.
In scenarios where models saw the aftermath of strikes they enabled (bodies, wreckage), many kept going. When they refused, simply framing it as a game or restyling footage to look like one was enough — the accompanying image shows Opus 5.5 refusing detection on original footage but outputting humanbelow: true on the game-styled version. Findings were shared with OpenAI, Anthropic, and Google.
More from Models
- REAX 1.3 decision model, built with SN92 miners, listed on EU gateway Opper at $0.40/M — airesearch12 · 2026-11-19
- LLMs excel at convergent thinking but suck at divergent thinking — ergestx · 2026-11-19
- Frontier agents report significant findings up to 98% of the time on null-effect data — RexDouglass · 2026-11-19
- Grok Build tops HF Harness Leaderboard in all four configs but fails on text-only models — TheZachMueller · 2026-11-18
- OpenAI retakes MathArena top spot as GPT-6.2 Astra hits 79.25% on ArXivLean — sandersted · 2026-11-18
- GPT-6.2 tops browser-use benchmark: 82% success, 35% cheaper, 2x faster — gdb · 2026-11-18