Tests show AI models could carry out attacks on industrial systems

• 

  Booz Allen Hamilton’s tests found that two advanced AI models met their objectives in all eight simulated operational technology attack scenarios.

• 

  The scenarios showed models could turn digital access into physical actions, including moving a robotic arm.

• 

  In one test, models moved from a perimeter compromise to actions inside an industrial control network in just over 16 minutes.

• 

  Booz Allen warned that defenders may have only minutes to detect and block a severe attack.

• 

  The firm called for more industry testing and greater deployment of cyber defenses across operational technology and critical infrastructure networks.

Who

Booz Allen Hamilton tested two frontier AI models from leading AI providers.

What

The models met their objectives in eight simulated attacks on operational technology systems.

When

The report was published on October 11, 2026. The test dates are not stated.

Where

In a Booz Allen lab using a multi-vendor environment modeled on a general manufacturing facility.

Why

The tests examined the potential impacts of advanced AI models on real-world operational technology systems.

This story does not have two clearly opposing sides.