The question for executives is no longer whether artificial intelligence will affect cybersecurity. It is whether their organization is still operating on assumptions from a pre-autonomy world.
Newly introduced frontier models from OpenAI and Anthropic threaten the capability to exploit software vulnerabilities at a scale no human team can match.
A new evaluation released Monday (April 13) by the U.K. Government’s AI Security Institute (AISI) conducted evaluations of Anthropic’s Claude Mythos Preview and found it to have successfully crossed into the early stages of operational cyber capability.
In simulated environments, the model did not simply execute isolated commands. It stitched together reconnaissance, exploitation, persistence and lateral movement into a coherent attack sequence. It adjusted its approach when steps failed, and maintained continuity across stages that traditionally required human oversight.
Historically, complex cyberattacks have been constrained by talent. Skilled operators are expensive, scarce, and often tied to a nation or well-funded criminal groups.
The latest models by the world’s biggest AI providers could represent a crucial cybersecurity inflection point. AI is no longer just a tool in the hands of an attacker; it is beginning to replicate aspects of the attacker itself.
Advertisement: Scroll to Continue
Read more: AI Is Cracking Open Banking Before Quantum Gets the Chance
For CFOs and CISOs alike, the implication is an increasingly stark one. Cyber risk is shifting from a targeted phenomenon to something closer to ambient exposure. Organizations are not just selected; they are continuously scanned, probed and tested by systems operating at scale.
The median enterprise, the one with uneven patching, over-permissioned accounts, and inconsistent configuration management, is now more accessible to multistep intrusion attempts that can be executed, or at least orchestrated, by AI systems.
Still, the most important takeaway from the Mythos evaluation is not that AI can already execute flawless cyberattacks. It cannot. As the U.K. report noted, the success rate is partial, the model’s capabilities are constrained, and its deployment remains controlled.
But systems that can plan and execute multistage intrusions, even inconsistently, represent a baseline that will improve. More compute, better orchestration, and tighter integration with external tools will incrementally close the gap between partial and reliable capability.
For CISOs, that means designing for a world where sophistication is no longer rare. For CFOs, it means recognizing cyber risk not as an occasional disruption, but as a persistent, evolving cost of doing business in a digitized economy.
The baseline has moved. The question is who will move with it.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
