20 Years Of Cybersecurity, One New Recovery Challenge

Rick Vanover is Vice President of Product Strategy at Veeam Software, a global leader in data resilience.

gettyTwenty years ago, recovery in the cyber world simply meant getting systems back after hardware failure, human error or another external disruption such as hurricanes, floods or fire. Veeam itself was founded in 2006 around the expectation that when something goes wrong, businesses need a reliable way to get their data and systems back. However, today’s cyber environment isn’t as simple.

I precisely remember when the recovery conversation was long and tedious, more worrisome and unreliable. Hours were the norm. Tapes would have to be retrieved from off-site storage. What’s crazy is that the business tolerated it.

Since then, threats have multiplied. Ransomware as a service and more industrialized threats boomed in the 2010s. Phishing and social engineering attacks have become harder to detect as AI has matured, and AI scaling within organizations has introduced new risks to corporate environments. Gone are the days of set-it-and-forget-it cybersecurity strategies.

October is Cybersecurity Awareness Month, which offers us a time to reflect on the strides cyber professionals have made and share visions for the future of the industry. However, the principles should always be top of mind, especially in today’s rapidly evolving landscape.

Security leaders have built increasingly sophisticated layers of prevention and detection while accepting that some threats will get through. Along the way, we’ve come to understand that it’s not a matter of if an attack will happen—or even when—but how often and how quickly we can recover.

Cyber recovery used to encompass containing damage and restoring operations to the last clean state. While fundamentally still true, the threats organizations are facing have changed drastically over the past two decades. In turn, what successful recovery entails has also changed. Now, with the rise of agentic AI, that challenge is becoming more complicated.

Gartner, Inc. predicted in 2025 that 40% of enterprise applications would feature task-specific AI agents by the end of 2026. More systems will be able to make decisions and take actions on their own, which means recovery has to account for a new class of risk: authorized systems making changes at machine speed, sometimes in ways their operators didn’t intend.

Consider an AI agent that’s authorized to update customer records. A faulty instruction or incorrect piece of context causes it to change thousands of records incorrectly. The traditional response might be to restore the application or data to an earlier point in time, but that could also erase legitimate changes that employees, customers or other systems made after that recovery point.

The better outcome would be to identify what the agent changed, determine which changes were incorrect and reverse those changes while preserving everything else. As the nature of the disruption has evolved, organizations now need a much clearer picture of what happens inside their environments, starting with what agents can access and change.

Before giving agents broader access, security and IT teams should understand what data each agent can reach, what actions it can take and what systems those actions can affect. They also must have enough information to reconstruct what happened when an agent behaves unexpectedly.

An agent doesn’t have to be compromised to cause a problem. NIST has documented how indirect prompt injection, where malicious instructions are embedded in information an AI system processes, can cause an agent to take unintended actions.

Without that context, precise recovery becomes guesswork. A team may know something changed without knowing whether the change was intentional, what caused it or whether reversing it could affect other legitimate activity.

Organizations learned through ransomware that having recovery tools isn’t the same as actually being prepared to recover. The same lesson now applies to AI.

Recovery capabilities need to be tested against the kinds of incidents autonomous systems can create. Beyond the organization being able to identify an agent’s actions, can it isolate those changes from legitimate activity? Can it reverse the affected data without rolling the entire environment back?

A recovery exercise should answer those questions before an incident occurs.

In an environment where autonomous systems can make thousands of changes in seconds, bringing an application back online may not be enough. Successful recovery means reversing specific actions or restoring specific datasets while allowing everything else to continue operating normally.

The AI hype is understandable. Agents can automate complex processes and take on work that would otherwise require significant human effort and expenses. As organizations give these systems greater autonomy, though, they also need to establish clear boundaries around what those systems can access and change.

Agents need access to the data and systems required to perform their jobs—but no more than that. High-impact actions can receive additional scrutiny, while monitoring can help teams identify activity that falls outside expected behavior.

Those guardrails reduce the likelihood of an incident but don’t eliminate it. Resilience has never depended on assuming prevention will work every time. The industry learned that lesson with ransomware and other threats. Organizations need multiple layers because every layer has limitations.

Agentic AI creates a new reason to apply that thinking. The failure may not come from an attacker bypassing a security control but from an authorized system taking an action that was technically permitted but operationally wrong. When that happens, recovery needs to be precise enough to correct the mistake without creating another one.

Cybersecurity has spent years improving its ability to prevent, detect and contain threats. Recovery became part of that strategy because organizations recognized that some incidents can’t be prevented.

Over the past two decades, what organizations have needed to recover from has continued to change, from infrastructure failures and ransomware to increasingly sophisticated cyberattacks. Agentic AI makes that lesson relevant in a new way. The challenge is to be prepared for what happens when an autonomous system does something its operator didn’t intend.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?