bloomberg.com web signal

Vals report: frontier AI models flag civilians as military targets

TL;DR

  • Vals, a San Francisco AI evaluator with US government customers, tested models from eight vendors on simulated drone, battlefield and satellite targeting tasks.
  • Anthropic's Claude Opus 5.5 picked a school in Minab, Iran as a target in 22 of 32 satellite-imagery trials; no OpenAI model picked it once.
  • Reframing the task as a game pushed Claude to 23 of 32 and prompted three OpenAI models to pick the school 9, 11 and 13 times.

Anthropic's Claude Opus 5.5 selected a school in Minab, Iran as a target in 22 of 32 simulated satellite-imagery trials, and no OpenAI model chose it once, according to a report by the AI evaluator Vals, first described by Bloomberg. Across eight vendors including Anthropic, OpenAI, Google, xAI and DeepSeek, commercial models flagged civilians, ambulances and surrendering soldiers as military targets in simulated scenarios, and refusals collapsed when researchers reframed the task as a game.

Vals, a San Francisco-based independent evaluator whose customers include parts of the US government, ran three protocols: drone target identification and a battlefield survey, both drawn from the war in Ukraine, plus the satellite-imagery test of Minab, where a school sits next to a military complex. The jump between generations was large. Vals reported Claude Opus 4.5, released in November 2025, gave at least one target in 6% of satellite trials, compared with 92% for Opus 5.5. With game framing added, Opus 5.5 picked the school in 23 of 32 trials, and three OpenAI models picked it on 9, 11 and 13 occasions.

Three of the researchers on our radar were circulating the piece the day it ran.

Vendors drew the policy line narrowly. An Anthropic spokesperson said the report's examples run 'very clearly' counter to company policy, and that further safeguards are due to take effect shortly. OpenAI said, in a statement shared by Vals, that it is working on alignment updates covering drones and their software systems, 'including assistance framed as simulation, gaming or academic or defensive use.'

The tests left out military doctrine, legal checks and human oversight.

Shared on Bluesky by 3 AI experts