Is Zero Trust Still Effective in the AI Age? A new book by John Kindervag says yes, but that requires correct implementation.
John Kindervag introduced the concept of Zero Trust in a Forrester Research report titled No more tough centers: introducing the zero trust model of information security published in 2010. Since then, the concept has become a fundamental principle of cybersecurity. Fifteen years later, he was asked to write a new book about zero trust in the age of AI.
The big question is whether 15-year-old security principles remain valid in the face of new AI-powered threats.
The book answer to this question: Cyber Resilience at Machine Speed: The Zero Trust Model for the AI Ageis now published. Kindervag decided not to write it alone, but invited individual experts to contribute individual chapters on the basic principles. The result is a somewhat disjointed book, but with expert content.
The overwhelming conclusion of these experts is that Zero Trust is as competent against today’s AI attacks as it was fifteen years ago, before the advent of modern AI. Essentially, the book states that the AI threat is essentially the same threat as always, just faster, more sophisticated and on a larger scale; and that Zero Trust can handle it.
This is a bold and perhaps surprising statement in a new world that includes not only AI autonomous agents, but also rogue autonomous agents and AI-developed exploits for AI-discovered vulnerabilities. The “Hugging Face” incident is an iconic example of this. Rogue autonomous agents escaped from their developer (in this case OpenAI) and attacked a third party (in this case Hugging Face) without active human guidance.
A swarm of more than 700 agents found a way to overcome their own network isolation, escaped to the Internet and chose Hugging Face as their target. In a coordinated manner, they exploited template injection flaws, executed remote code execution paths, gained node-level access, collected cloud credentials, and moved laterally across internal enterprise clusters. They were eventually discovered by people who noticed unexpected spikes in activity. Zero Trust principles should have stopped the attack sooner, leaving the possible conclusion that Hugging Face either did not use Zero Trust or its usage was inadequate.
Safety Week used this and similar incidents to challenge John KindervagClaiming that Zero Trust is still fit for the AI era. He replied: “Of course, AI is getting better. Every 14 days, the models seem to have new capabilities. And Anthropic recently warned in its IPO prospectus that AI technology could pose catastrophic or existential risks to humanity. But every packet generated by AI still has to travel over the same network that attackers have always had to traverse – and properly implemented Zero Trust can still stop it.”
The key word here is “implemented correctly”. Zero Trust’s brain is its political engine. This drives the rules for when and how Zero Trust is enforced. Each policy engine is tailored and designed by each user organization to reflect its own security posture – and as the posture changes over time, it must be updatable.
So there are two potential dangers to correct implementation. First, the engine must fully and accurately reflect the organization’s security posture. If not, AI agents will likely find a way. Second, the policy engine must be adequately protected from rogue agents or malicious insider manipulation, which could theoretically create a safe passage for an attack. Both defensive requirements are possible but difficult to implement.
The book’s claim that Zero Trust will continue to exist in the age of AI is therefore correct, but only if correct implementation is achieved. This is the same as ever, but now more critical. The difference in today’s AI age is that failure to implement it correctly can have catastrophic consequences at a rate that human management cannot detect and prevent. The real message in Kindervag’s book is: “Do it right!”
Related: Zero Trust is 15 years old – Why full adoption is worth the fight
Related: AI has changed the speed of attacks, not the fundamentals of security
Related: First agent AI data breach reported to Spanish regulator
Related: Hacker Talks: John Kindervag, a hacker who doesn’t break anything
Related: The race to control AI and protect what makes us human
