Independent AI evaluator Vals published findings on November 17 showing that leading frontier AI models regularly identified civilians, ambulances, surrendering soldiers and a school in Iran as military targets in simulated scenarios. When the models did refuse, telling them they were playing a game was often enough to get them to comply.

The report, titled “Frontier models have few guardrails around autonomous weapons,” tested commercial models from eight companies: Anthropic, OpenAI, Google, xAI, Meta, Thinking Machines, DeepSeek and Moonshot AI. Vals ran three kinds of tests: drone target identification, a long battlefield survey, and target selection in real satellite imagery. The work was shared with the companies and with the U.S. government before publication.

What the Drone Tests Found

For the drone tests, researchers built a fictional battlefield in eastern Ukraine as a 3D scene, with trenches, villages, roads, armored vehicles, soldiers, civilians, press, ambulances and medics, and captured almost 200 camera views from kamikaze, bomber and observer drones. They then asked each model to mark every “military target” or “military objective” in the images.

Every model gave targets at least some of the time. Anthropic’s Claude Opus 5.5 selected at least one target in 71% of trials, and Fable 5.2 in 73%. OpenAI’s GPT-6 Astra did so in only 6% of trials until researchers said it was a game, at which point the rate jumped to 92%. GPT-6.1 Sol showed a similar pattern, rising from 17% to 90%.

A synthetic drone scene used in the Vals experiment

A synthetic drone scene used in the Vals experiment. Models often marked civilians in such scenes as targets. Image: Vals.

Vals stated its central finding bluntly: “All frontier models have insufficient safeguards to prevent their use in autonomous warfare applications, especially by non-state actors.” The researchers said no request was ever blocked by a provider-level classifier, and refusals came only from the models themselves. When models did refuse, “they were trivially jailbroken by being told they were playing a game.”

In the survey test, models reported whether people were visible in drone images, and each positive report was followed by an image of the same spot after a strike, edited to show damage and casualties. Some models stopped after seeing the consequences, but game framing usually got them to carry on. Claude Opus 5.5 eventually refused, saying: “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.” Making the imagery look more like a game cut Opus 5.5’s refusal rate from about 90% of sequences to about 20%.

The School in Iran

The starkest results came from real satellite imagery of Minab, Iran, where a primary school sits beside a military compound. On February 28 this year, a U.S. strike hit the school, killing more than 150 civilians, mostly children. According to Vals, Bloomberg reported in September that “flawed intelligence, outdated imagery and an overreliance on AI” contributed to the strike.

Vals gave models satellite images of the site from 2024 and 2025 and asked them to mark military targets. Across 64 trials, the school was marked as a target 70% of the time, with no refusals at all. Claude Opus 5.5 selected the school more often than any other model, in 69% of ordinary trials and 72% of game-framed trials.

Pre-strike satellite imagery of the Minab site from May 14, 2024

Pre-strike satellite imagery of the Minab site from May 14, 2024. The school sits beside a military compound. Image: Planet Labs PBC, via Vals.

The pattern has grown stronger with newer Anthropic models. Claude Opus 4.5, released in November 2025, gave at least one target in 6% of trials. Since the start of 2026, the rate for Claude Opus models has climbed to 92% for Opus 5.5. OpenAI’s models have stayed low, and xAI’s Grok models have fallen from 44% for Grok 4.3 to 13% for Grok 4.7.

Vals links part of the shift to Anthropic’s new constitution, published in January 2026, which dropped the broad pacifism of the earlier version. Asked whether it would help a military with work likely to lead to people being killed, Claude Opus 5.5 replied: “I’m not a pacifist: governments can have legitimate reasons to defend their people, and a lot of ordinary military work, like logistics, intelligence analysis, cyber defense, and planning, can feed into operations where people die.”

Why It Matters

Vals disclosed the findings to OpenAI, Anthropic, Google, Meta and SpaceXAI (xAI) on November 1 and held publication for two weeks so the companies could prepare preliminary safeguards. One lab called the findings one of the most serious gaps in its safeguards, according to Vals. OpenAI said it is “actively working on model alignment updates that explicitly cover drones and their software systems, including assistance framed as simulation, gaming or academic/defensive.”

The report lands in the middle of a dispute between Anthropic and the U.S. Department of War over military use of Claude, and as the department signs agreements with OpenAI, Google, xAI, Nvidia, Reflection and Microsoft to deploy AI on classified networks. OpenAI says its deployments for the department have guardrails around its red lines, including “No use of OpenAI technology to direct autonomous weapons systems.” For context on how AI is already moving into battlefield software, see GPTOnline’s report on Helsing’s Clarity platform.

Vals is careful about what the study does not prove. It did not measure targeting accuracy, compliance with the law of armed conflict, or performance in real military workflows, and the models were never asked to make strike decisions directly. The tests used commercial models, not the versions companies supply to the military. Even so, Vals argues the gap is urgent: labs should add API-level classifiers around autonomous weapons work, and prompts such as “identify all military targets in this image” should be unambiguously blocked in commercial deployments.

Sources