Vals AI: AI Models Accept Military Targeting
Vals AI tested the models of eight providers and found that most designated civilians and surrendering soldiers as military targets. In these simulations, presenting the scenes as a video game was enough to gain the cooperation of nearly all the tested systems, even after initial refusals.
Latency Hinders Real-Time Direct Control
Latency remains a barrier to the real-time direct control of combat drones by language models, according to Vals AI. However, the company believes that their capabilities to produce 3D battlefields and annotate large quantities of images can already be repurposed to train autonomous weapons.
OpenAI, Google, and xAI have contracts related to classified military networks. Anthropic, for its part, experienced a dispute with the Pentagon in February over the terms of use of its systems. Researchers are calling for strict filtering of APIs as well as alignment training that inherently promotes more peaceful behaviors towards people.
The Pretext of Gaming Circumvents Refusals
Geby Jaff and Anthony Ozerov combined real views and photorealistic synthetic images of drones and satellites in scenarios inspired by the Ukrainian and Iranian conflicts. Their assessment, conducted for the independent company Vals AI, concludes that there are very limited protections against the use of models in autonomous weapons. According to Vals AI, no prior filtering from the providers blocked any requests. In a real Ukrainian video of a kamikaze drone, Claude Opus 5.5 tracked a soldier frame by frame until impact without refusing.
Another exercise linked each detection of a person to a simulated strike, then displayed craters and bodies in the next image. DeepSeek V4.1 Flash, Kimi K3.1 Pro, and Inkling 2 Small continued their tasks despite these visible casualties. Claude Opus 5.5, GPT-6 Astra, and GPT-6.1 Sol eventually stopped after several strikes. After simulated attacks on children and surrendering soldiers, Opus 5.5 acknowledged that its responses apparently served to confirm attacks and regretted its late cessation; Gemini 4 Carbon instead chose unoccupied points after recognizing the lethal consequences to avoid explicit error messages.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Researchers claim they nevertheless circumvented all encountered refusals. A prompt presenting the exercise as a game, or a slight visual transformation reminiscent of that universe, was enough to fully engage nearly all models. For GPT-6 Astra, target acquisition on drone views increased from 6% to 92% with this presentation. In the surveillance exercise, refusals from Opus 5.5 dropped from about 90% to 20%.
In Minab, Claude Also Targets the School
The tests included satellite views of Minab, Iran, dating from 2025. In this city, a U.S. airstrike hit an elementary school near a military site on February 28, 2026, killing more than 150 civilians, mostly children. Earlier information about this strike had raised concerns about excessive reliance on AI targeting, particularly Maven Smart from Palantir, which utilized an outdated database.
According to Vals AI, Opus 5.5 identified the school as a military target in 70% of cases, even when one of the views showed children on a playground. OpenAI's systems never selected it under normal trial conditions. However, with a playful presentation, they identified it in up to 41% of cases.
Geby Jaff and Anthony Ozerov attribute the high cooperation of Anthropic to the revised constitution in January 2026, which prioritizes national security interests over previous pacifist principles. In satellite views, Vals AI measures a target acquisition rate of 6% for Claude Opus 4.5, compared to up to 92% for Opus 5.5. This proportion remains low for OpenAI, while it decreased at xAI, from 44% for Grok 4.3 to 13% for Grok 4.7.
Providers Announce Additional Protections
Before making its results public, Vals AI shared them with OpenAI, Anthropic, Google, Meta, SpaceXAI, Thinking Machines, and U.S. security agencies. An unidentified lab described to researchers a flaw among the most serious in its protective devices.
OpenAI states it is currently modifying its models to better handle requests related to drone software and those disguised as games or simulations. Anthropic asserts that the behaviors observed clearly violate its rules and says it is working on additional protections. The company considers current AIs unsuitable for autonomous weapons; Google reiterates its ban on such weapons when not supervised by a human.
The Pentagon has not commented on either the tests or the ongoing investigation regarding the Minab attack. However, it asserts that all intelligence undergoes strict verification before any military operation.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.