Companies are moving quickly to deploy agentic and autonomous systems, with more than 60% of organizations expecting to use AI agents within the next two years, according to Gartner research. The push is being driven by the potential to automate complex workflows, increase productivity, and move from AI-generated insights to autonomous action. Adoption speed and time to deployment are becoming competitive differentiators in their own right.
If you don’t catch the right train at the right time, whatever your company is working to deliver with AI technology may be table stakes or, worse yet, yesterday’s news, upon its arrival.
As AI systems gain greater access to enterprise data and become capable of making autonomous decisions and taking actions, enterprises must understand and address the increased risk this presents.
Without proper safeguards, automation can scale mistakes, expose sensitive data, and propagate issues more quickly. This is not to say that the answer is to slow adoption; it means that to be successful, companies need to build governance, data protection and cyber resilience into enterprise AI strategies from the beginning.
Changing the game, compounding the risk
Everybody talks about how automation helps with scale. But the same AI capabilities that enable enterprises to automate at scale also ratchet up their risk and that of their stakeholders.
AI makes it easier for attackers, fraudsters and other bad actors to replicate exploits and do more damage. Automation that is flawed, due to model drift or other often-unseen issues, may amplify mistakes at scale. That’s a real problem when AI exists in enterprise IT systems, but it can prove even more catastrophic in the physical world.
The concern is no longer science fiction. Remember the nefarious droids in Star Wars? The idea of a droid army isn’t as far-fetched when you take into account the number of humanoid robots poised to exceed 1 billion by 2050, according to Morgan Stanley Research. That means the risk of physical AI being ultimately corrupted by the dark side is a legitimate concern.
The fallout from compromised physical AI systems could come in many forms. An attack on an autonomous vehicle could cause a car or truck to decelerate when it should accelerate, potentially resulting in a crash. Tampering with sorting and packaging robots could cause them to put pills in containers labeled as other medicines, creating a public health emergency.
However, as the recent MIT article published by the World Economic Forum that highlighted these scenarios explained, “today’s allocation of funding does not reflect the scale of the risks.”
Growing threats attract government and regulatory attention
As ransomware and other forms of cyberattack have evolved into highly sophisticated ecosystems powered by new AI tools, regulators are taking notice. The UK has proposed prohibiting public sector bodies and critical national infrastructure from making ransomware payouts.
Meanwhile, government leaders and regulators in the U.S. and abroad are emphasizing resilience. The White House recently unveiled its America First Resilience Strategy. The Securities and Exchange Commission requires public companies to follow rules for cybersecurity risk management, strategy, governance and incidents. The European Union’s Digital Operational Resilience Act (DORA) has applied in full since January 2025, with no transition period, and financial entities are now under active regulatory supervision to meet its requirements. And the Bank of England’s Prudential Regulation Authority is requiring organizations to demonstrate their recovery capabilities.
Cyber resilience is taking on new importance amid increasing AI risk and emerging regulations.
Enabling explainability, visibility and governance
With enterprises and governments, including sovereign cloud operators, increasingly concerned about controlling the data on which AI relies, many organizations are bringing pre-trained large language models into controlled enterprise environments, where sensitive data can be more tightly governed, fine-tuning those models with new and proprietary information, and deploying them for consumption and inference.
But even if you bring an LLM into your enterprise, how can you tell whether you’re using suitable data to which you have the rights? And how can you be sure you are compliant?
Some organizations are building this kind of visibility directly into their data infrastructure. They track the data stored and retrieved through vector databases and trace its ownership over time, which is essentially a flight recorder for how a model was trained and fine-tuned. This visibility enables verifying whether data sources are valid and auditing yourself to ensure that your AI systems are governed in a way that aligns with company policies and government regulations.
If an AI system acts autonomously, influences a loan decision, makes a prediction or recommends a business action, those answers and actions must be accurate. But organizations also need to know why the system reached that answer, what data influenced it, and how that data is governed and used. That requires explainability, visibility and data governance.
Extending cybersecurity and fighting AI with AI
Growing AI risks and regulatory requirements provide CISOs, CIOs, security leaders and enterprise decision-makers with new incentives to ensure security is an intrinsic capability. By extending cybersecurity closer to the data and infrastructure layer, they can benefit from the ability to detect anomalous activity quickly and contain potential issues before they propagate.
Using AI across protection, detection and recovery strengthens enterprise defenses. Imagine a bad actor trying to access or corrupt a volume of data. Having data infrastructure that can detect that anomaly and either limit the actor’s interaction with that data, take the data offline, stop replication of that data or take snapshots of the data can significantly limit your risk.
Because bad actors are using all the tools at their disposal to penetrate enterprise defenses, organizations must assume that they have been, or will soon be, compromised. To limit the damage, enterprises should embrace zero-trust architectures, which never trust and always verify. That principle should be extended to AI agents as well.
Expediting data restoration with modern data protection
Data sits at the heart of modern businesses. When data is deleted or held hostage, it can bring business processes to a halt. That can result in business disruption and, in turn, legal and regulatory problems, lost customers and revenue and reputational damage. The severity of these problems depends on what data was deleted and how long it takes to recover lost data.
Restoring deleted data following a cyber incident or a disaster such as an earthquake, flood or fire used to be a long, manual process. But it’s now possible to recover data in minutes. This requires a clean copy of the data. To make that possible, locate mission-critical data in an air-gapped part of your environment, or ensure you have the capability to reconstruct your application or business process environment using data you can prove has not been infected.
FHNW University of Applied Sciences and Arts Northwestern Switzerland recently adopted such a solution to improve its business continuity and enforce backup best practices. In the process, it achieved 3x to 4x acceleration of tape backups, increased cloud data backup-to-disk retention from one month to five years and reduced its cyber risk with immutable backups.
Data infrastructure that is built with resilience in mind, including immutable snapshots and anomaly detection, can make a difference when an attack hits. Hitachi Vantara’s Virtual Storage Platform One (VSP One), for example, has a 100% data availability guarantee with clean data recovery in as little as 30 seconds.
Shifting the conversation from speeds and feeds to the data that drives AI
To date, much of the AI conversation has revolved around model capabilities and GPU speed. Now, as enterprises move to scale AI, they need to focus on data. That includes understanding where data came from, where it goes and should not go, and how to protect and recover it.
The organizations that scale agentic AI successfully will be those that pair greater autonomy with equally strong governance, visibility and resilience.
