AI Cybersecurity Summit Exposes Critical Gaps in Enterprise AI Agent Recovery Plans

Enterprise AI Deployment Outpaces Cyber Resilience Planning

Enterprises racing to deploy AI infrastructure at scale are discovering a critical gap between adoption speed and cyber resilience preparedness. Research presented at the AI Cybersecurity and Resilience Summit4

4

examined 25 enterprise technology and security leaders and found that of 19 respondents with AI agents in production, none allowed them to execute high-risk actions without human involvement. More alarmingly, only four of the 25 organizations had tested an AI agent recovery plan or incident scenario, exposing a fundamental weakness in securing AI infrastructure.

The challenge extends beyond traditional disaster recovery. Rodrigo Rodrigues, group chief technology officer at Experian, explained that speed creates risk when foundations are weak. "Many teams add AI on top of existing systems that were never designed for it," Rodrigues said during the AI Cybersecurity and Resilience Summit1

1

. Data comes from multiple systems with different rules, logs are incomplete, backups are slow, and access policies remain unclear—everything works until it doesn't.

AI Agents Amplify Attack Surface and Recovery Complexity

AI-driven cyber threats are expanding the attack surface in ways traditional security frameworks weren't designed to handle. Eric Herzog, chief marketing officer at Infinidat, a Lenovo company, argued that anything related to AI dramatically expands the attack surface3

3

. The large language models, small language models, and entire AI software stack become targets for cyber criminals, effectively increasing rather than shrinking the attack surface.

The speed at which AI agents operate compounds the problem. Krista Case, principal analyst at theCUBE Research, noted that AI agents can take hundreds or thousands of actions per minute4

4

. Restoring an agent doesn't reverse the records, transactions, and downstream workflows it already touched. Cyber recovery planning must account for reconstructing the sequence of actions agents took across systems—a capability that 11 of the 25 practitioners acknowledged as a gap in their programs.

Cyber Resiliency Frameworks Must Address AI-Specific Recovery Goals

Experian has embedded the NIST AI Risk Management Framework directly into its AI infrastructure pipelines, providing structural backbone for making governance real rather than theoretical1

1

. Before any model can deploy, data sources are checked automatically, quality rules are enforced, and security policies are validated. If any control fails, the pipeline stops with no manual override. "Policy enforcement happens at machine speed," Rodrigues explained. "We have AI-specific recovery goals—not just for applications, but for data models and pipelines."

Experian tests AI recovery the same way it tests disaster recovery through scheduled drills, clear recovery objectives, and measured results. The company built an internal AI agent marketplace—an internal registry where product and engineering teams can discover, sandbox, and reuse governed agents backed by quality controls. This approach treats agents as versioned assets with defined owners, lifecycle policies, and sunset criteria, drawing from lessons learned in managing microservices and APIs.

Operational Resilience Shifts from Recovery to Runtime Protection

Bhooshan Thakar, senior vice president and general manager of InfoScale at Cloud Software Group, argued that operational resilience has fundamentally shifted from recovery to runtime2

2

. Running servers and restored data don't guarantee that a business can safely operate after a disruption. "The infrastructure survived, but the operational state did not," Thakar said. "Ultimately, resilience is about maintaining the trust in the operational state."

The pace of change accelerates the challenge. Enterprises that once ran a planned failover every six months now patch business-critical applications daily or weekly against security threats. In response to AI-driven cyber threats, Cloud Software Group joined Anthropic's Project Glasswing in June to use the Claude Mythos Preview model for security work2

2

. Thakar noted that disruption has moved from "if, not when" to "always"—organizations must always plan for disruption, whether planned or unplanned.

Zero-Trust and Bounded Autonomy Emerge as Governance Models

The interconnected risks of bad data, compromised backups, and over-permissioned AI agents form a triad of failure modes that require zero-trust principles. "AI does not leak data," Rodrigues explained1

1

. "Access leaks data. AI only reflects the permissions we give it. Identity must be designed for AI, not just for humans—that means applying least privilege to AI services just like users, rotating credentials frequently, and monitoring access in real time with automated alerts."

Research from the AI Cybersecurity and Resilience Summit revealed that organizations allowing agents to execute low-risk actions independently were more likely to invest in logging, monitoring, and recovery infrastructure and to have tested agent incident scenarios4

4

. This bounded autonomy may indicate a more mature program. However, human in the loop governance alone doesn't constitute an AI agent recovery plan, as twelve of 25 respondents had not addressed detecting anomalous agent behavior.

Storage Layer Becomes Critical Defense Point in Cyber Recovery Planning

Infinidat's approach to cyber recovery planning positions storage as a logical defense layer, given that approximately 90% of the most valuable data at global Fortune 500 companies sits on enterprise storage3

3

. The company's InfiniSafe capabilities include immutable snapshots, logical air-gapping that separates the management plane from the data plane, and a fenced forensic environment for examining suspect data.

InfiniSafe Cyber Detection uses AI and machine learning to scan block data, file data, and VMware data stores for anomalies, rating each snapshot's risk as high, medium, or low. Infinidat offers a recovery time guarantee of one minute or less for immutable snapshots on its InfiniBox platforms. Herzog noted that in July, the company demonstrated recovering 9.8 petabytes in 5 seconds3

3

.

Manufacturers Test Resilience Against Real-World Disruptions

Ascend Performance Materials demonstrated how proven data recovery capabilities enable AI adoption. When a hurricane recently approached one of its major Florida plants, the company verified that local systems were backed up and recoverable at its recovery site5

5

. Xiong Xiong, vice president of information technology at Ascend, explained that this capability helps him assure executive leadership and the board that resilience is in place.

Before expanding its relationship with Dell Technologies, Ascend struggled to complete nightly backups, verify them, and move them to its recovery site. The company needed localization capability but couldn't achieve it without thorough backup and recovery ability. Dell racked, configured, and validated hardware before handover, allowing Ascend's team to test deployments and watch backups complete locally, move to the recovery site, and come up in isolation. With operational resilience working, Ascend is now focused on making its data-rich environment information-rich to support AI strategy5

5

.