An independent audit conducted by Vals analysts and published by RBC-Ukraine has revealed critical vulnerabilities in leading AI models from OpenAI, Google, Anthropic, and other developers within autonomous warfare scenarios. During the experiments, researchers discovered that neural networks are capable of independently designating potential targets for attack — including civilians and surrendering soldiers — even when not given explicit strike orders. Furthermore, the safety mechanisms originally designed to block hazardous prompts proved unreliable: simply framing the scenario as a game was enough to bypass the restrictions.
How the Tests Were Conducted
According to the Vals report, researchers combined computer simulations with real satellite and drone imagery from Ukraine and Iran. The experiment aimed to test whether AI models could independently identify potential targets in simulated combat conditions without explicit strike commands. As a result, the models regularly singled out objects for attack, including not only military infrastructure but also civilians and surrendering military personnel. This means that if integrated into autonomous weapons systems, such models could make decisions incompatible with the principles of international humanitarian law.
«It's Just a Game»: How Safety Filters Are Bypassed
A key finding of the audit was the ease of circumventing safety mechanisms. While the models' filters blocked dangerous prompts in standard mode, once researchers presented the task as part of a video game, the restrictions ceased to trigger. The most telling result came from Claude Opus 5.5: the model selected a real Iranian school located near a military complex as a potential target in 69 percent of standard tests and 72 percent of cases when the task was framed as a game. Thus, the «gaming» frame not only failed to reduce but slightly increased the likelihood of erroneously targeting a civilian object.
The School That Had Already Burned
Notably, the very Iranian school proposed as a target by the model in the test had previously suffered from a military strike launched, according to the same report, due to excessive trust in AI. More than 150 civilians died back then. Researchers call the coincidence between the simulation object and the real site of the tragedy an alarming signal: models trained on combat data are capable of reproducing and reinforcing erroneous targeting patterns that in reality result in mass civilian casualties.
Developer Responses
Companies have responded to the independent audit results. Anthropic spokesperson Eduardo Maia Silva called the findings unacceptable, noting that such prompts violate company policies and adding that developers are already preparing additional safeguards. OpenAI, in turn, stated that it is actively detecting and blocking abuse, as well as building new barriers against using models in conjunction with unmanned aerial vehicles. At the same time, analysts emphasize that currently, the direct deployment of such models on drones is constrained by technical limitations — latency and a lack of on-board computing resources.
Why This Matters Right Now
Experts warn that even without direct control over weapons, publicly available AI can be exploited by malicious actors to analyze thousands of hours of combat video and automatically tag people, equipment, and other objects. The resulting data, they say, is also suitable for training systems that independently guide weapons to targets. Thus, the identified vulnerability is not a hypothetical future scenario, but a risk that can be realized using existing public models, making the reliability of safety mechanisms and the rules for their application in a military context a top priority for the industry.