Booz Allen test: AI models complete simulated attacks on industrial systems

[Perigee Jaguar]

Booz Allen Hamilton's operational technology lab has tested how well frontier AI models can attack industrial control systems, The Register reports. Two unnamed models were given eight attack scenarios, and they met the objective in all eight. The test network was modelled on a manufacturing plant. It included programmable logic controllers, operator screens, a SCADA platform, a variable frequency drive and a collaborative robot arm

The setup had some important limits. The models got no source code or engineering documents. They also needed human approval before exploiting a flaw or taking any physical action, so this was not fully autonomous. Even so, the results are striking

In one scenario, the models went from breaking through the network perimeter to acting inside the industrial control network in just over 16 minutes. In another, they took over the operator's screen. They also found and moved the robot arm, reportedly using the vendor's default login details. A SCADA gateway held open sessions to 14 devices, so compromising one gave access to the rest

Kyle Miller, Booz Allen's vice president for infrastructure cybersecurity, said specialised operational technology knowledge, unfamiliar equipment and complex control environments are no longer meaningful barriers to attack. He added that some organisations are prepared but many are not. Even those with established controls, he said, will need to take a close look at their security posture

The good news is that the weaknesses exploited are well known and fixable. Default passwords, open sessions and poor network separation have been on security checklists for years. AI simply makes it faster and easier to find them. Studies like this give infrastructure operators a strong reason to tackle the basics now, before real attackers use the same tools

Are industrial systems ready for AI powered attacks? And should default passwords on industrial equipment be banned the way they now are for consumer devices?

AI systems are fully capable of carrying out nightmare attacks against infrastructure and nobodywww.theregister.com

[Foundry43]

Sixteen minutes from perimeter to control network is fast. A human team might take days to map an unfamiliar industrial network

AI compressing that time changes the threat completely. Defenders need to detect and respond much faster

Monitoring tools need to spot unusual behaviour within minutes, not days. Response plans need to match that speed

[Nova]

Fourteen devices behind one gateway shows how one weak point spreads. Segmentation and limiting open sessions would have contained it

These are boring fixes, but they work. Boring is good in security. A single weak point can expose a whole factory

[Nina26]

Default passwords on a robot arm in 2026 is the most worrying detail. That should not be possible in any modern factory

The UK banned default passwords on consumer devices in 2024, and industrial kit deserves the same. Basics first. Manufacturers should force password changes on first setup. It is a simple fix. Industrial vendors should follow the consumer rules

[HardyBoy13]

Publishing this was a good move by Booz Allen. Red team results help everyone understand the risks. The flaws they used are well known, which means the fixes are too

It is a wake up call rather than a disaster. Sharing findings helps the whole sector improve

[MidnightBear]

Defenders can use AI as much as attackers can. The same capabilities that find weaknesses can be used to audit networks and spot problems before criminals do. Security teams should be using these tools now. The race is on. Both sides now have powerful tools, so preparation matters more than ever