Rogue AI can wipe your database in seconds. A guide to accelerating safe AI at scale.

“AI agents have become your most productive, communicative, collaborative internal employees. But they bring it with speed. And unfortunately it has some very interesting consequences,” said David Allott, Field CISO for APJ at Veeam Software, at a closed-door Data and AI Trust Leadership Roundtable held in April in Hong Kong. Veeam Software is the number one ranked data protection software provider globally, according to IDC’s 2025 annual software tracker, growing faster than all competitors in the top five in the category.

The data trust crisis

Customers are encountering new and serious challenges around their data. When neither can leaders fully trust the data feeding their AI systems nor the actions those systems take, the entire AI investment risks becoming toxic. Excessive permissions, toxic combinations of sensitive data and over-privileged agents, and the inability to recover cleanly from mistakes are all amplifying risk at scale.

Veeam’s Data Trust & Resilience 2026 Report shows many organisations are struggling to maintain visibility and control over data flows across apps, clouds, and third party services. Globally, 43% say AI adoption is outpacing their ability to secure data and models, while 40% say security policies have not yet been updated to address AI specific risks.

To address these emerging risks, Veeam acquired Securiti AI and merged its data resilience expertise with advanced data security and governance capabilities to deliver a unified approach offering visibility, control and recovery at AI speed.

Controlling rogue AI: Detect, Protect and Undo

Powered by the firm’s proprietary Data Command Graph capability, Veeam’s Agent Commander solution is a relational intelligence engine that maps connections between sensitive data, user identities, AI models and autonomous agents across hybrid environments.

Tim Stead, Technical Director for Data and AI Security in APAC at Securiti (now part of Veeam), explained the platform’s core strength. He called on enterprises to equip themselves with the ability to detect AI risks in context, protect systems at runtime and undo agent mistakes with surgical precision – capabilities he said Agent Commander now delivers.

The Agent Commander platform works across three layers of defence. 

• The first detects shadow AI agents and models, providing a comprehensive inventory with hosting details, entitlements and activity timelines. 
• The second protects data through policy enforcement, least-privilege adjustments, data sanitisation before ingestion into AI systems, and runtime controls such as prompt injection prevention. 
• The third enables precise recovery – surgical restoration of only the files or records affected by an agent mistake, drawing directly on Veeam’s backup infrastructure.

Stead illustrated the platform’s value with real customer examples. For instance, a major US telecom provider analysed 69 petabytes of mostly unstructured data and removed a large volume of redundant, obsolete and trivial information. This reduced storage costs, shrank the attack surface and delivered much cleaner datasets for its AI initiatives.

In another case, a large US bank paused its Google Gemini Enterprise rollout after discovering excessive data exposure, then used the Data Command Graph to classify information and correct access rights before safely proceeding.

Stead also demonstrated how “toxic combinations” turn manageable risks into major incidents. For example, an AI Copilot with access to acquisition plans, over-permissioned data sources and inherited user rights can quickly trigger sensitive data leakage.

“We are helping businesses stay focused on data security, backup and resilience, while adding advanced features to address the new challenges created by AI,” said Joseph Chan, Regional Vice President for Greater China and Mongolia at Veeam. “Agent Commander takes data resilience and security to another level preparing businesses for the AI era. It provides the foundation layer to successfully deploy AI, unleashing the power of data and AI in a safe and secure manner.”

Allott reinforced the governance imperative. Boards are increasingly aware of their personal liability for AI-related failures, while regulators continue to tighten expectations.

 “With trusted data, we can adopt AI agents at scale with trust. It’s removing those concerns that have really been blocking, slowing down the pace of adoption,” said Chan.

(Sponsored Content)