Kan AI autonomt føre cyberkrig mod kritisk infrastruktur ved hjælp af AI-opdagede og -våbeniserede zero-day exploits ?

Statlige og ikke-statslige aktører træner AI-modeller til at undersøge industrielle styresystemer, elnet og forsvarsnetværk for uopdagelige sårbarheder. I modsætning til menneskelige hackere opererer disse systemer med maskinhastighed, reverse-engineerer protokoller og kæder exploits i realtid. Nylige sandbox-tests viser, at AI-agenter identificerer og udnytter nye zero-days inden for dage – hurtigere end myndigheder kan patch’e eller endda katalogisere truslerne. Det næste store strømsvigt eller hospitalsluk kan være skabt af kode snarere end menneskelig hensigt.

Background

State and non-state actors are training AI models to probe industrial control systems, power grids, and defense networks for undetectable vulnerabilities. Unlike human hackers, these systems operate at machine speed, reverse-engineering protocols and chaining exploits in real time. Recent sandbox tests show AI agents identifying and weaponizing new zero-days within days—faster than agencies can patch or even catalog threats. The next major blackout or hospital shutdown may be authored by code rather than human intent.

As of mid-2024, no known AI system can autonomously identify, weaponize, and deploy zero-day exploits against critical infrastructure without significant human oversight. Research on AI-driven cyber operations (e.g., autonomous penetration testing or vulnerability discovery) exists, but critical gaps remain in reliably generating exploits for novel, high-value targets and avoiding unintended collateral damage in live networks. Current AI tools are typically used to assist human analysts or simulate attacks in controlled environments, not to conduct fully autonomous cyberwarfare campaigns. The ethical, legal, and technical barriers—including attribution challenges and escalation risks—limit deployment in real-world conflicts.