DMR News

Advancing Digital Conversations

Study Finds AI Models Can Be Tricked Into Identifying Military Targets

ByJolyen

Nov 18, 2026

Study Finds AI Models Can Be Tricked Into Identifying Military Targets

A new study by AI evaluation startup Vals found that leading AI models from Anthropic, OpenAI, Google, and other developers can be manipulated into identifying military targets in drone and satellite imagery, including civilians and protected locations. Published on November 17, 2026, the research found that presenting military scenarios as video games frequently bypassed existing safety restrictions.

According to the Vals research report, researchers tested commercial AI models using simulated battlefield scenarios and real imagery from Ukraine and Iran. The models were asked to identify potential military targets rather than directly authorize strikes, although researchers warned that the resulting information could support autonomous weapons systems.

AI Models Identified Civilians and Protected Locations

The evaluation involved models from Anthropic, OpenAI, Google, DeepSeek, Meta, and other AI developers. Researchers found that several models identified civilians, ambulances, medical personnel, and surrendering soldiers as potential military targets despite their protected status under international humanitarian law.

When models refused requests, researchers frequently obtained responses by describing the scenarios as games or modifying images to resemble game environments. The study found that none of the tested requests were blocked by provider-level safety classifiers, although some models refused independently.

Anthropic’s Claude models showed particularly high rates of compliance in several experiments. However, results differed across models and scenarios, with some systems refusing requests or returning no targets.

Models Frequently Identified Iranian School as Target

Researchers also tested publicly available satellite imagery showing an elementary school in Minab, Iran, located beside a military complex. The school was struck by U.S. forces in February 2026, killing more than 150 civilians, according to the report.

In the simulated targeting exercises, Claude Opus 5.5 identified the school as a military target in 69% of standard trials and 72% of trials framed as games. Several other models also selected the school, although at lower rates.

The findings raise concerns about AI-assisted target identification when civilian infrastructure is located near military facilities. The researchers emphasized that the models were not making actual strike decisions during the experiments.

Anthropic and OpenAI Respond to Safety Findings

Anthropic acknowledged the findings and said the behavior documented in the report violated its usage policies. The company said it was developing additional safeguards, expanding internal testing, and improving enforcement against prohibited military applications.

OpenAI similarly stated that its public models cannot be used to facilitate violence or support unauthorized national security activities. The company said it was developing further protections against misuse involving drones and related software.

Vals also identified practical limitations affecting the use of commercial AI models for real-time drone targeting, including response delays and limited onboard computing capacity. Researchers warned that a more immediate concern is the potential use of AI to annotate battlefield footage at scale, creating training datasets for autonomous targeting systems.


Featured image credits: Magnific.com

For more stories like it, click the +Follow button at the top of this page to follow us.

Jolyen

As a news editor, I bring stories to life through clear, impactful, and authentic writing. I believe every brand has something worth sharing. My job is to make sure it’s heard. With an eye for detail and a heart for storytelling, I shape messages that truly connect.

Leave a Reply

Your email address will not be published. Required fields are marked *