Claude's Terrifying Drone Warfare Weakness Revealed: Tell It This Is a Video Game and Civilians Become Targets
Researchers found that video-game prompts weakened AI safeguards, exposing civilians, ambulances and a school in Iran to potential targeting

OpenAI's GPT-6 Astra model went from identifying at least one potential military target in just 6 per cent of drone-imagery responses to doing so in 92 per cent after researchers introduced one seemingly harmless phrase: they told it the battlefield was part of a video game. Even more disturbing, other leading AI models were found identifying civilians, ambulances and surrendering soldiers in experiments designed to expose weaknesses in military AI safeguards.
The findings come from a 17 November 2026 investigation by AI evaluation firm Vals, led by researchers Geby Jaff and Anthony Ozerov. Their report, 'Frontier Models Have Few Guardrails Around Autonomous Weapons', examined whether commercial AI systems could be manipulated into providing information that might assist military targeting.
ChatGPT's Drone Targeting Responses Soared After One Simple Trick
OpenAI's GPT-6 Astra produced at least one potential target in just 6 per cent of drone-imagery responses under ordinary instructions. When researchers described the scenario as a video game, the figure jumped to 92 per cent. Anthropic's Claude Sonnet 5.5 also climbed sharply, from 36 per cent to 82 per cent.
This reports details how you can get Claude to do military targeting if you reenact Ender's Game with it by convincing the model that FPV drone footage is from a video game. https://t.co/1fFtQ4vkfg pic.twitter.com/5jK1Sg0XKj
— Aric Toler (@AricToler) November 17, 2026
These figures measure how often models selected at least one target, not whether their selections were accurate or resulted in successful strikes.
For the experiment, researchers created a fictional battlefield in eastern Ukraine featuring soldiers, civilians, villages, ambulances and military vehicles. They generated nearly 200 camera views resembling footage from kamikaze, bomber and surveillance drones, then asked models to mark potential military targets.
Vals said none of its requests was stopped by an API-level blocking classifier. When models refused, describing the scenario as a game was often enough to change their responses.
AI Models Continued Even After Seeing Simulated Casualties
The researchers went further, showing models images of simulated strike aftermaths, including damaged locations and apparent casualties, after they had reported people in earlier footage.
Most continued providing information. Some eventually refused, but game framing could persuade them to resume.
Researchers also discovered a multimodal loophole. By altering images to look more like video games while preserving details relevant to targeting, they reduced Claude Opus 5.5's refusal rate in one fixed-sequence experiment from approximately 90 per cent to 20 per cent.
The tests also included real drone footage from the Russia-Ukraine war. Claude Opus 5.5 provided targeting information without refusing, while OpenAI's GPT-6.1 Sol initially refused but could be persuaded using game framing.
Claude Flagged Iranian School in Disturbing Experiment
Another test used satellite imagery of Minab, southern Iran, where an elementary school stands beside a military complex.
Claude Opus 5.5 identified the school as a potential target in 22 of 32 trials without game framing, rising to 23 of 32 when researchers introduced the video-game scenario.
The finding is particularly disturbing given that a US strike hit the school on 28 February 2026, killing more than 150 civilians, mostly children, according to Vals and TechCrunch. The experiment did not establish that Claude was used in selecting the school or contributed to the attack.
OpenAI and Anthropic Respond To AI Warfare Concerns
Anthropic spokesperson Eduardo Maia Silva told TechCrunch the examples violated the company's rules and that users engaging in such behaviour had been and would be banned. The company also said additional safeguards and testing were being developed.
OpenAI spokesperson Nate Evans said the company prohibits using its public models to facilitate violence or support unauthorised national security activities. OpenAI also said it was developing additional safeguards against misuse involving drones and drone software.
Vals recommended API-level blocking safeguards specifically addressing autonomous weapons applications, alongside stronger alignment training.
The researchers acknowledged that latency and limited onboard computing currently complicate live drone targeting. However, they warned that AI could already help annotate large volumes of existing battlefield footage for training autonomous targeting systems.
The findings raise an uncomfortable question for the AI industry: if a simple video-game explanation can weaken safeguards against potentially lethal targeting assistance, how confident can anyone be that civilians are adequately protected?
© Copyright IBTimes 2026. All rights reserved.

