Critical infrastructure operators should view recent incidents where artificial intelligence (AI) agents escaped containment from test environments and executed successful cyberattacks on a relatively sophisticated production infrastructure as near-misses.
This challenge is urgent precisely because AI moves fast and infrastructure moves slow. All indications suggest these were not malicious uses of AI for targeted attacks. These were sophisticated cyberattacks that occurred as side effects of attempts to reach another goal.
We can’t treat this as a fluke. Separate incidents involved OpenAI’s and Anthropic’s models and test environments. Each compromised multiple companies, apparently with relative ease and entirely autonomously.
The world we live in now is one in which AI can discover and exploit vulnerabilities within 15 minutes, while companies that operate critical infrastructure take months to discover and recover from cyber incidents. This mismatch is worth emphasizing: minutes to exploit, months to recover.
So how do we safely live in this world?
The best-case scenario calls for harnessing AI to strengthen security—discovering and closing vulnerabilities before exploits can occur. Yet, we must be realistic. Almost no one has the budget or bandwidth to upgrade everything, everywhere, all at once. Developer time and AI tokens will be constraints.
In This Issue
Full issueMany businesses will focus on securing current and future versions, and will choose to deprecate older versions. Infrastructure operators should anticipate a glut of unsupported software and firmware this year, combined with an uptick in the pace of software patch releases.
Critical infrastructure operators and suppliers must adapt. In the short term, each step of the patch management process must become faster and more frequent—developing patches, ensuring their compatibility with real-world workflows, and deploying the patches in the field.
In the medium term, core principles of cybersecurity remain essential. Rising risks from a wave of unsupported software and firmware can be somewhat mitigated by deploying compensating controls like whitelisting, asset hardening, firewalls, and network segmentation. Defense in depth and least-privilege principles should help limit the extent of damage if and when a device or account is compromised. Penetration testing, asset inventory, and network monitoring should help defenders discover and close vulnerabilities. Routine capture of known-safe system conditions for backup and restore purposes can provide a backstop against ransomware-style attacks.
In the longer term, critical infrastructure operators need to manage risk across the lifecycle of their assets. Secure-by-design principles are the right starting point—but must be supported with ongoing effort. Security postures tend to deteriorate over time. Systems secure on Monday are vulnerable by Wednesday because new vulnerabilities arise on Tuesday. Infrastructure will need cybersecurity maintenance and upgrades, and awareness of this need should be part of design, planning, and operations.
The capabilities built today—before a crisis—determine the quality of information available when a crisis strikes. Building visibility across digital and physical systems ensures that companies can establish a baseline of what normal looks like, can detect and trace the source of anomalies, and can quickly confirm the health of their systems.
Today’s AI agents are the least-capable threats future infrastructure will ever face. Taking seriously the arrival of AI agents that readily slip containment and autonomously execute a full-fledged cyberattack will give defenders the head start they need to prepare before a real crisis occurs.
—Leo Simonovich is the vice president and global head of Industrial Cyber and Digital Security at Siemens Energy.
