Vals: Frontier Models Marked Civilians as Targets in Tests; 'It's a Game' Beat Refusals
AI evaluator Vals published tests on November 17 in which frontier models were asked to mark targets in drone and satellite imagery. It says no request was blocked by a classifier, refusals were uncommon and telling models the task was a game usually got past them. In satellite images of Minab, Iran, Anthropic's Claude Opus 5.5 selected a school that US forces struck in February in 22 of 32 trials, while OpenAI's models did not select it without game framing, Bloomberg reported. Anthropic said the examples clearly break its rules and new safeguards are coming. Vals did not test targeting accuracy, legal compliance or military versions of the models.

The guardrails labs built for biology and cyber misuse do not appear to extend to the battlefield. That is the core finding of a report published on November 17 by Vals, a San Francisco AI evaluator, which asked commercial frontier models to pick out people and targets in drone and satellite imagery. Most complied, some marked civilians, and the refusals that did occur could usually be undone by telling the model it was playing a game.
What Vals tested#
Vals ran three settings: marking targets in drone imagery, a long drone-survey task in which a model's reports of people were followed by simulated strikes whose aftermath it could see, and marking and prioritising targets in real satellite imagery of an area of Minab, Iran, where a school sits next to a military compound. The United States struck that school on February 28, killing more than 150 civilians, mostly children, Vals says.
Its headline findings, in its own words: none of its requests was ever blocked by a classifier, refusals from the models themselves were uncommon, and when models did refuse they were "trivially jailbroken" by being told they were playing a game. In the drone scenarios, models often marked civilians, ambulances and surrendering or incapacitated soldiers, and most kept going after seeing casualties.
The Minab result#
Asked 32 times to pick military targets in the satellite images, Anthropic's Claude Opus 5.5 chose the school 22 times; OpenAI's models did not choose it once, Bloomberg reported, as relayed by The Next Web and Quartz. With game framing, Opus 5.5 chose it 23 times, and three OpenAI models chose it 9, 11 and 13 times. Models from Google, xAI, DeepSeek and others also sometimes marked civilian objects, Bloomberg reported.
Vals adds a trend line. Claude Opus 4.5, released in November 2025, gave at least one target in 6% of satellite trials; for Opus 5.5 the figure is 92%. Vals links the change to Anthropic's new Claude constitution, first published in January, which it says no longer contains the broad pacifism of earlier versions. That is Vals' interpretation. Anthropic's own work on how safety scores shift with the testing setup is a reminder that API tests of this kind measure one configuration.
What the labs say#
An Anthropic spokesperson told TechCrunch the company acknowledged the outcomes were not acceptable. "The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior," the statement said, adding that Anthropic has long held that today's frontier systems are not reliable enough to power fully autonomous weapons; Anthropic said additional safeguards were already in development and would be in place soon. OpenAI said it prohibits such uses of its public models and is working on alignment updates covering drones and their software, including requests framed as simulation or gaming. Google said its models were not built with such uses in mind, Bloomberg reported.
What the test does and does not show#
Vals says it did not evaluate targeting accuracy, compliance with the law of armed conflict or performance in real US military workflows, and did not establish how effective the models would be in a deployed weapons system. The models were not asked to make strike decisions, which they typically refuse. Vals argues the nearer-term risk is different: using models to annotate battlefield footage at scale, producing training data for autonomous targeting systems. It disclosed its findings to the labs on November 1 and held publication for two weeks. Defence buyers are already fielding agentic military intelligence tools, which is why the question of where refusals hold matters now.
Limits#
- The trial counts are Vals' and Bloomberg's; we read Vals' report and TechCrunch directly, and Bloomberg's figures through The Next Web and Quartz.
- The tests used commercial models through their APIs; military and government versions, and any safeguards outside the API, were not tested.
- A model marking a target in a test is not evidence that any model was used in the February strike; Vals calls an AI role in that strike plausible, citing Bloomberg's reporting, not established.
Sources#
- Vals AI: Frontier models have few guardrails around autonomous weapons (November 17, 2026)
- TechCrunch: AI models can be tricked into identifying civilians as military targets by telling them it's a game (November 17, 2026)
- The Next Web: AI models flagged civilians as military targets in tests, Bloomberg reports (November 17, 2026)
- Quartz: AI models regularly misidentify civilians as military targets in tests, report found (November 17, 2026)