AI agents are quickly moving from experimental tools into systems that can support customers, trigger workflows and make decisions across enterprise environments. For CX leaders, that raises a practical question of identity in accessing frontier AI models: how do you verify who, or what, is using them when agent behavior can shift in real time? Anthropic’s decision to shut down Fable underlined the challenge, as there was no simple way to comply with U.S. government restrictions on non-U.S. nationals accessing the model.
In this CX Today interview, Nicole Willing spoke with Geoffrey Mattson, CEO of SecureAuth, about the identity infrastructure enterprises may need as agentic AI becomes more capable. The identity restrictions on Anthropic’s Mythos and Fable AI models expose a bigger issue, Mattson said: powerful models can be hard to govern if providers cannot reliably verify who is using them and why.
Mattson argued that this could push the market toward Know Your Customer for AI, especially for models with possible dual-use applications. While financial services has long had obligations to confirm customer identity, AI providers may face similar pressure to understand the intent behind model access. That matters because traditional login details offer little assurance when autonomous systems are acting through application programming interfaces and delegated workflows.
“Some of these models are getting more and more powerful. As Anthropic pointed out, a lot of people don’t realize there is absolutely no way to make a jailbreak-free model. In fact, a lot of people think it’s relatively trivial to get LLMs to do what they’re not supposed to do. These things are impossible to control, and they are potentially capable of doing real damage, from creating cyber weapons and bioweapons to facilitating other forms of criminal activity.”
For customer experience teams, the warning is immediate. AI support agents are increasingly tied to service outcomes, customer trust, compliance, and brand risk. Mattson pointed to cases where customer support bots have given incorrect guidance, leaving businesses responsible for the result. As agents spawn subagents, accountability can become even harder to trace.
That is why identity needs to evolve from a static login check into something more dynamic. Mattson explains that enterprises must understand who owns an agent, what systems it can access, and when a human should approve higher-risk actions. Guardrails alone are not enough if models can behave unpredictably or be manipulated through prompts.
Mattson’s message is not to slow innovation. It is to separate agents from what they can access, apply stronger identity controls and keep humans in the loop for higher-risk decisions. For enterprises exploring AI in CX, the competitive edge may come from making security simple enough that teams can innovate without leaving accountability behind.
