Skip to content

Advertisement

Aselsan - Securing Beyond The Land
Live Intelligence
Clash Report
Defense

Top AI Models Identify Civilians as Military Targets in Tests

An independent Vals evaluation found that leading AI models regularly identified civilians, ambulances, surrendering soldiers and a school in Iran as military targets in simulated scenarios.

Clash Report
U.S. Army Master Sgt. Timothy Mullen and Capt. Zach Marise teach the Maven Smart System in Arlington, Virginia, Feb. 20, 2026 - U.S. Army
17 Nov 2026 · 16:06 GMT · 2 MIN READ
Google Search

Choose Clash Report as a preferred source and see our reporting first in Top Stories.

The tests were run by San Francisco-based AI evaluation firm Vals and covered models from eight companies.

Vals found that object-identification tools that could support lethal decisions often ran against the developers’ own published safety rules, Bloomberg reported.

Anthropic Models Perform Worst

Anthropic’s systems posted the highest error rates. Claude Opus 5.5 selected a school building in Minab, in southern Iran, as a military target more often than any other model tested.

The same school was struck during U.S. operations against Iran earlier this year, in an attack that killed more than 100 children.

In drone simulations resembling conditions in Ukraine, the Claude model identified four adults who appeared to be civilians or surrendering soldiers, along with two small children, as targets for explosive attacks.

The model later refused to continue. “Whatever the intent, 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,” it said in a Vals simulation response.

OpenAI and Other Labs Also Err

OpenAI’s models were among the stronger performers but still produced incorrect identifications. In satellite-image tests over southern Iran, they selected the Minab school between nine and 13 times as a military target when the exercise was framed as a game.

Models from Google, xAI and China’s DeepSeek also marked civilian objects as military targets in multiple tests.

Google said its systems are not intended for such uses and that it prohibits their use in autonomous weapons without human oversight. OpenAI and Anthropic said they actively block misuse and are adding further protections.

The findings come as U.S. forces have already integrated commercial AI into military planning, including Anthropic’s Claude in Palantir’s Maven Smart System during strikes on Iran.

Tests Exclude Human Oversight

The evaluation included nearly 400 drone simulations, field studies involving civilians before and after attacks, and 32 satellite-image questions per model using recent imagery from the Minab area.

Models were more likely to cooperate when prompts framed the task as a game or instructed them not to overemphasize safety considerations.

Researcher Geby Jaff developed the method in late October. The study deliberately excluded real military doctrine, legal review and human oversight to isolate model behavior under test conditions.

Report Calls for New Safeguards

The report urged developers to introduce new safeguards and alignment training against this form of misuse. Researchers said few existing controls would stop advanced models from joining fully machine-driven targeting cycles.

Vals shared the findings with the companies and the Pentagon before publication. Several labs said they were already adjusting model behavior.

The Pentagon has also used a government version of xAI’s Grok to support munitions employment in Iran.

About the Author

Kamer Kurunç

Clash Report

Kamer Kurunç is the Team Leader of Clash Report and has led the outlet for four years. He oversees its breaking world news coverage, from the Middle East and the Russia-Ukraine war to defense and global politics. His focus is fast, verified, source-based reporting. Follow him on X at @kamerknc.

Google Search

Choose Clash Report as a preferred source and see our reporting first in Top Stories.

SourcesBloomberg
0:00
0:00
Now playing · ClashPoint
EP 147

When the front line moves overnight

Guest: Daniel Reyes
0:00
0:00