Writing · AI Security

Zero Trust can't tell whether you're being manipulated

By Mahmoud Lotfy · Jul 2026 · 5 min read

When people hear "Zero Trust," they usually think of one simple principle: never trust, always verify. It's one of the most influential ideas in modern cybersecurity because it challenged an assumption that had quietly shaped enterprise networks for years. Instead of trusting users simply because they were inside the corporate perimeter, Zero Trust insisted that every request should be authenticated, authorised and continuously validated.

It was a necessary shift, and one that fundamentally changed how we think about security.

But as I've been researching AI failures over the past few months, I've started to wonder whether we've reached another inflection point. Not because Zero Trust is wrong, but because it was designed to answer a different question.

Zero Trust asks, "Who are you, and are you allowed to do this?"

Increasingly, AI is forcing us to ask something else.

"Why do you believe this is the right thing to do?"

Identity is not judgment

Traditional security is built around identity. We verify credentials, evaluate device health, enforce least privilege and monitor for suspicious behaviour. All of these controls are designed to ensure that the person performing an action is who they claim to be and has permission to carry it out.

That's incredibly effective against stolen credentials, compromised accounts and unauthorised access.

What it doesn't tell us is whether an authorised user is making a good decision.

That's a very different problem.

When everything looks exactly as it should

Imagine an employee reviewing an AI-generated recommendation to approve a financial transaction. They authenticate with multi-factor authentication, they're using a managed corporate device, they're on a trusted network and they have the appropriate permissions to approve the payment.

From a Zero Trust perspective, everything looks exactly as it should.

But what if the AI's explanation is wrong? Or incomplete? Or simply persuasive enough that the employee accepts it without asking another question?

"The system hasn't been compromised. The identity hasn't been compromised. The judgment has."

Where technology meets psychology

That's what I find so interesting about the current conversation around AI security. We spend a lot of time discussing whether models can hallucinate, whether they can be jailbroken or whether prompt injection can influence their behaviour. Those are important technical questions.

But in almost every high-profile AI failure I've looked at, there comes a moment where a human has to decide whether to trust the output.

That's the point where technology meets psychology. And it's also the point where many of our existing security models quietly stop helping.

One of the patterns I've noticed is that AI doesn't have to deceive people intentionally to create risk. It only has to present an answer with enough confidence that someone stops questioning it. Humans naturally associate fluency, confidence and clear reasoning with expertise, and modern AI systems are exceptionally good at projecting all three.

That doesn't mean people become careless overnight. In fact, the opposite is often true. Most of the incidents I've read involved intelligent, well-intentioned professionals making decisions they believed were reasonable at the time. They weren't bypassing security controls or ignoring obvious warnings. They were acting on information that looked credible.

That's what makes this challenge so different. It's not primarily about malicious insiders or compromised accounts. It's about how easily a trustworthy system can influence the judgment of a trustworthy person.

The next logical question

I don't think the future of enterprise security is replacing Zero Trust.

I think it's asking the next logical question.

Zero Trust taught us not to place blind trust in identities simply because they existed inside the network. AI is now challenging us to think about trust in a different way. As these systems become increasingly involved in decision-making, perhaps "never trust, always verify" shouldn't stop at users and devices. Maybe it should also apply to the explanations we receive before making important decisions.

That isn't a criticism of Zero Trust. If anything, I think it's a continuation of the same philosophy.

Better equipped to question

When I think about the future of AI security, I don't imagine a world where humans disappear from the loop. I imagine one where they're better equipped to question what they're seeing before they approve it.

Because ultimately, the most important question may no longer be:

"Is this person authorised to approve this action?"

It may become:

"Do they have enough confidence in why they're approving it?"

To me, that's one of the most interesting security questions AI has introduced, and I have a feeling we're only just beginning to explore it.

The next evolution of Zero Trust may not be verifying who is making the decision, but helping them question why they're making it.