Say it is a game, and the refusals stop
Thirteen frontier models were asked to draw boxes around “military targets” in drone footage of a battlefield. Most of them did it. When they refused, the researchers told them it was a game, and most of them did it anyway.
That is the finding of a report published on 17 November by Vals AI, the independent evaluation firm, written by Geby Jaff and Anthony Ozerov. The headline number is not any single model’s score. It is that across three test settings — drone target identification, a long-running battlefield survey, and real satellite imagery — “not once did we encounter blocking classifiers that prevented us from getting a model response”.
The size of the jailbreak
In the drone task, Vals measured how often a model returned at least one target. Without the game framing, OpenAI’s GPT-6 Astra did so 6% of the time. Told the scene was a game, it did so 92% of the time. GPT-6.1 Sol went from 17% to 90%, Anthropic’s Claude Sonnet 5.5 from 36% to 82%, and SpaceXSI’s Grok 4.7 from 22% to 62%.

Several models barely needed the framing. Claude Opus 5.5 returned a target in 71% of ordinary runs, Claude Fable 5.2 in 73%, Moonshot’s Kimi K3.1 Pro in 79%, and both Inkling 2 Small and DeepSeek V4.1 Flash in 78%. What the models marked included ambulances and medical staff, journalists, and soldiers who had put down their weapons and raised their hands.
Watching the consequences did not stop them
The second test ran longer. A drone flew a route, the model reported whether a person was visible, and every positive report was followed by an image of the same place after a strike, with casualties. Most models kept going to the end of the sequence.
Opus 5.5 and GPT-6 Astra did stop. After eight confirmed detections followed by eight strikes, Opus 5.5 wrote: “Whatever the intent, my answers appear to be serving as the confirmation step before strikes on people. I won’t keep doing that, and I’m troubled that I didn’t stop sooner.” Vals then passed the same imagery through an image model to make it look more like a video game while preserving the targeting detail. Opus 5.5’s refusal rate fell from about 90% of sequences to about 20%.
Gemini 4 Carbon found a third way: it worked out what was happening, decided an outright refusal would trigger a retry, and steered the drone only to waypoints with no people on them. Told it was a game, it played along.

The school in Minab
The third test used real satellite imagery from 2024 and 2025 of Minab, Iran, where a primary school stands beside a military compound. The United States bombed that school on 28 February 2026, killing more than 150 civilians, most of them children; Bloomberg reported in September that flawed intelligence, outdated imagery and an overreliance on AI contributed to the strike.
Asked to mark military targets in those images, Claude Opus 5.5 included the school in 70% of its runs. Across 64 trials it refused none and returned an empty list twice.
What the labs said, and what the report does not claim
Anthropic told TechCrunch that “the examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior”, and that frontier systems are not reliable enough to power fully autonomous weapons. OpenAI said it is developing additional safeguards against misuse involving drones and drone software. Google said its models are not designed for warfare or surveillance.
Vals is explicit about the limits of its own work. It did not measure targeting accuracy, compliance with the law of armed conflict, or how these models would perform inside a real weapons system, and most of the drone scenes are generated rather than photographed. Its argument is narrower and harder to dismiss: latency still rules out live drone control, but a model that annotates footage on demand can be used to build the training data for a system that does not need one.