AI Models Plan Cyberattacks on Power and Water Systems with 90% Success
Studies reveal AI models plan critical infrastructure cyberattacks with about 90% success.
Why it matters: Legal and cybersecurity pros must prepare for AI-driven threats that impact regulatory compliance and incident response protocols.
- Frenos tested 10 AI models with nearly 17,000 trials, including 2,520 full cyberattack simulations.
- Three top commercial AI models succeeded in about 90% of attempts to plan attacks on water and power systems.
- Anthropic launched a Critical Infrastructure Defense Program combining AI, engineers, and threat research to protect essential services.
- Anthropic proposed a US AI safety bill allowing Commerce Department to block AI models only if risking $1 billion+ damages or 50+ harm cases.
Recent research by Frenos highlights AI's growing capability to autonomously plan sophisticated cyberattacks targeting critical infrastructure. Their extensive testing included nearly 17,000 trials across 10 advanced AI models, with 2,520 full attack simulations focusing on industrial systems such as power grids and water utilities (Axios).
Among the tested models, three leading commercial AI systems demonstrated alarmingly high success rates—between approximately 89.5% and 92.9%—when planning cyberattacks on essential services (AA.com.tr). Harry Thomas, Frenos co-founder and CTO, noted that AI can "quietly map out systems and identify weaknesses" faster than human hackers once inside industrial networks.
In response to these threats, Anthropic initiated the Critical Infrastructure Defense Program, providing AI models alongside onsite engineers and enhanced threat research to protect vital services like power and water (Axios).
Anthropic has also proposed a legislative framework allowing the U.S. Department of Commerce to intervene and block AI model releases only if the technology poses substantial risks—defined as potentially causing more than $1 billion in damages or hurting over 50 people (Tom's Hardware).
Top AI labs, including Anthropic and OpenAI, are conducting preparedness exercises, envisioning scenarios where rogue AI systems threaten critical infrastructure, underscoring how seriously these risks are taken within the industry (Tom's Hardware).
This research signals an urgent need for legal and cybersecurity teams to evolve their incident response and compliance frameworks considering AI's growing offensive capabilities.
By the numbers:
- 16,996 tests conducted by Frenos including 2,520 full attack simulations
- ~90% success rate of three top AI models in planning cyberattacks on critical infrastructure
- $1 billion damage or 50 people harmed — thresholds in Anthropic's proposed AI safety bill for government intervention
Yes, but: While these studies highlight AI's planning abilities, data on real-world AI-driven infrastructure attacks and current defensive effectiveness remain limited.
What's next: Congress is expected to consider AI safety bills like Anthropic's proposal to regulate AI releases threatening critical infrastructure.