Should companies plan to switch AI models?

As AI models compete for users amidst uncertain government regulation, experts say that the ability to replace AI systems is crucial for maintaining business operations.

In June, Anthropic temporarily disabled its most advanced frontier models, Fable 5 and Mythos 5, after the Trump administration ordered the company to suspend access to those models for all foreign nationals. The administration’s move reportedly stemmed from concerns over the potential for those models to be jailbroken and misused, particularly in cyberattacks.

Victor Wieczorek, SVP for offensive security at GuidePoint Security, said the temporary shutdown raised the question of whether organizations and their IT teams should plan for disruptions to their AI setups. For example, a company that previously relied on a single model for its AI work might consider a mix of frontier, legacy, and open-weight models, or even a balance between cloud-based and on-prem AI.

“This injected a lot of uncertainty, and what you have to do with uncertainty is take a step back and plan,” Wieczorek said. “So, now we’re getting a lot of people who are planning better than what they ever did.”

Handling uncertainty, safely. Joanna Chen, CISO for Dashlane, said that the smartest way to grapple with uncertainty around frontier models is to try and build infrastructure where models can be switched out if necessary. Applying principles such as isolating data in advance can help in the event of a transition.

“You will always need isolated environments,” Chen said. “You will always need to understand what data is flowing where, to who. You will need to understand where your secrets are, and are they being ingested by the AI.”

Chen said that, in the context of cybersecurity, it might be prudent for IT pros to keep “any kind of really sensitive data away from AI right now, especially there’s some things that should never touch AI,” such as encryption keys, passwords, or API tokens. Making sure that data doesn’t simply pass into a bucket of datasets being used to improve models is “very simplistic.”

There’s also a need for human-based oversight.

“The same principles apply, of least privilege, where you can segregate [or] isolate systems, when you can have detection and monitoring over them, make sure all the access is tracked and identifiable and attributed to whoever is responsible for that access,” Chen said. “When things go wrong, no AI agent is going to be held accountable, it’s still going to be a person. Who is going to be that person in your organization?”

Some advice. Wieczorek said, with government interference and the related uncertainty, he recommends professionals take a moment to “step back and plan” for how their organization can more efficiently deploy and customize tech to meet their needs.

“We’re going to get folks who are paying attention now, who are looking across a vast field of options, and able to now be choosy about what they want and customize the right solution for themselves,” Wieczorek said.