Cyber Resilience in the AI Era: Familiar Challenges, Higher Stakes

In my last post, I laid out the definition of cyber resilience I’ve used for years: if your business processes rely on IT systems and something disrupts those IT systems, through malice or plain error, can you get back to a known good state so the business can persist or resume?

Every word of that still holds with agentic AI. Though AI makes every part of that definition harder to actually satisfy.

That’s why Data Protection Matters (DPM) surveyed hundreds of IT leaders this year, on what we call AI-Enhanced Resilience. We wanted to know where AI is helping resilience and how organizations are ensuring the resilience of AI workloads themselves.

The definition above serves as a test an organization can use to determine whether it’s actually prepared. If that test gets harder to pass and nobody notices that the bar has been raised, organizations end up believing they’re resilient when they’re not, and the first place that gap shows up is usually in the middle of an incident, not before it.

Where AI raises the bar

If your business depends on a system, you’re depending on what its vendor claims. A good percentage of the IT leaders we surveyed said that some vendors are doing more marketing than practical implementation when it comes to AI in their backup, DR, and cyber tools. That gap matters more than it used to.

It’s not entirely the vendors’ fault. Every vendor has to talk about their AI-based capabilities today, just like every vendor has to frame their backup tool within a cyber resilience narrative. In both cases, it is what IT leaders are seeking, but sometimes the outcomes these tools promise aren’t as easy to achieve as the initial narrative or vision might have suggested. DPM’s earlier Organizational Resilience research last year also revealed some naïve optimism gaps between what senior leaders assumed was achievable and what their own implementers believed. To be fair, these aren’t easy problems to solve.

A new kind of disruption, and a harder recovery to prove

It used to be easier to identify what to protect against: an attacker, an honest mistake, a natural disaster. All look wrong while they’re happening, or immediately after. An attacker leaves tracks or a ransom note. Humans have those ‘uh oh’ moments. A weather event leaves no doubt.

AI introduces a new case that doesn’t announce itself the same way: a tool, acting on your behalf, doing exactly what it was built to do, and getting it wrong in milliseconds. Nothing about the action looks suspicious, because nothing malfunctioned.

I’ve had this happen to me directly, an AI assistant “fixing” something I’d already fixed by hand, with no way to undo it. Minor frustration at the scale of a slide deck. Magnify that to a sales database or a codebase, and the stakes change entirely. Restoring data doesn’t recall the customer email an agent already sent or the refund it already issued; that’s a continuity problem, not just a backup problem.

Most of what’s happening here is a hyper-acceleration of problems we already knew how to solve, just moving too fast for the old process to catch. If you’re being attacked or an agent is making errant changes with valid credentials at agentic speed and you’re waiting for a human to decide before you react, you’re already behind. Agentic threats require an agentic-speed response.

If you’re not testing, you don’t have a plan, you have a hope; that was true before AI and it’s still true now. What’s changed is how much harder the “known good state” is to define when the thing that changed your data was acting on its own judgment, not following a person’s process. There’s no clean ‘before’ to point to, since your own legitimate changes and the agent’s changes are stacked on top of each other in the same timeline.

Many of the organizations DPM surveyed were running modern data protection tools. Yet there is a lack of confidence that if AI started making changes, they could actually identify and unwind them.

Getting back to basics

I recently sat down with Rob Callaghan, Principal Solutions Marketing Manager at Wasabi, for part two of our conversation on cyber resilience in the age of AI. Even in the midst of so much modernization, we talked about one key recommendation for closing the gap between what AI promises and what organizations can actually trust it to do. We agreed that organizations must get back to the basics.

The problem of someone deleting data they shouldn’t be able to delete is not new, whether that meant burning paper records or wiping a drive on the way out the door. The solutions that work aren’t new either: multiple controls, approvals, or keys, so no single person can act alone, and a copy no one can touch, kept somewhere a compromised account can’t reach even if it tries. That same policy works for the paper records of yesteryear, for digital data and configurations today, and yes, even for agentic actors.

What’s different now is granularity and speed. Automate detection and containment so you’re not waiting on a person to notice. Keep human approval on anything irreversible. Scope what each agent can touch, rather than giving it broad access by default. And keep enough attributable versions to unwind one agent’s changes without losing everyone else’s work. Then ask any vendor claiming AI-centric capabilities within their solution to demo each of those scenarios.

Rob and I go much deeper on exactly that in the full conversation. Join us, and we’ll cover:

• Why threats moving at machine speed require faster detection and response

• Where organizations lack confidence in recovering from unintended AI changes

• What to consider when evaluating AI capabilities in your data protection tools

• How protected copies, multi-user approvals, and tested recovery plans support resilience as AI use grows