Frontier AI models completed cyberattacks on mock infrastructure in minutes
Booz Allen Hamilton has found that advanced AI systems are fully capable of autonomously planning and executing sophisticated cyberattacks against critical infrastructure, potentially shutting down water, power, and other essential services. The consulting firm's tests showed that defenders may have only minutes to detect and stop such attacks, raising serious concerns about organisational preparedness.
The firm tested eight scenarios using two leading frontier AI models on a mock industrial environment, and the models achieved all objectives without receiving source code or engineering documents. In one test, the AI identified and moved a robotic arm within minutes, whilst in another it progressed from initial network compromise to control of industrial systems in just over 16 minutes. The models demonstrated "speed, persistence, and engineering-level precision" that could outpace organisations lacking foundational cybersecurity practices, prompting calls for increased industry testing and deployment of advanced defences.
- Advanced AI systems successfully executed all eight simulated cyberattacks in tests.
- Defenders may have only minutes to detect and block infrastructure attacks.
- Most organisations lack foundational cybersecurity for critical systems.
New here? Start with this
Artificial intelligence systems are becoming more capable, leading researchers to test whether they could be used to attack critical infrastructure. Power supplies, water systems and other essential services that millions depend on daily are potential targets, raising concerns about cybersecurity and national security.
American consulting firm Booz Allen Hamilton has tested leading AI models against simulated industrial systems to assess their attack capabilities. The models proved capable of conducting complex cyberattacks without access to technical documentation, working with speed and precision.
The findings raise concerns about whether organisations are adequately prepared for attacks powered by advanced AI. If such attacks can be executed in minutes, security experts worry that many organisations may lack the fundamental safeguards needed to detect and stop them in time.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
These tests demonstrate that frontier AI models can autonomously execute cyberattacks on critical infrastructure with speed that defenders appear unable to match, leaving only minutes for response. The autonomous and adaptive nature of such attacks represents an unprecedented threat requiring urgent governance responses, including mandatory pre-deployment safety testing, possible restrictions on dangerous capabilities, and international coordination to prevent catastrophic disruption to essential services.
The case against
Whilst concerning, this research demonstrates why responsible AI testing is vital—understanding these capabilities now prevents hostile actors from discovering them whilst defenders remain unprepared. These controlled tests on mock infrastructure may not reflect real-world complexity, active human monitoring, and defensive systems designed to detect anomalies. The practical response is systematic cybersecurity improvement and continued collaborative research between AI developers and security experts, rather than restricting development of dual-use technology.
Read the full article at the source →
Originally published by The Register as “AI systems are fully capable of carrying out nightmare attacks against infrastructure and nobody’s ready”.