Cohesity launches backup and rollback for AI agents

Cohesity introduces backup for AI agent state

Data security firm Cohesity has launched Agent Resilience, a capability within its Data Cloud platform that backs up the memory and configuration of enterprise AI agents. The feature enables organizations to roll back agents to a trusted state when they malfunction or are compromised, addressing what the company describes as a critical gap in agent governance.

The announcement came at Cohesity's Catalyst conference, where the company also outlined plans to automate its cyber resilience framework using agentic workflows. Agent Resilience is currently available to select customers, with general availability expected by year-end.

The agent recovery problem

As AI agents gain traction inside enterprises, they increasingly operate with privileged credentials and query sensitive databases without human review of each action. While observability tools can detect when an agent behaves unexpectedly, Cohesity argues these tools cannot undo changes an agent has made to data or systems.

"Detection can tell you an AI agent went off course," said Vasu Murthy, Cohesity's chief product officer. "It cannot undo the changes."

Most organizations do not protect agent state today, according to Cohesity. A prompt injection attack can compromise an agent that appears normal on the surface, causing it to act on untrusted information. Without a clean recovery point, the only option may be to wipe and rebuild the agent entirely.

Agent Resilience protects both memory and configuration using the same immutable backups and clean-room recovery Cohesity applies to other workloads. When something goes wrong, administrators select a recovery point from before the suspicious behavior began, and only affected memory or settings are restored. The service also identifies and protects the databases, file systems, and vector stores each agent accesses, and generates topology maps showing what systems and data an agent can reach.

At launch, the service supports only agents built on Amazon Web Services' Bedrock platform, including Bedrock AgentCore and Bedrock Agents. Support for Microsoft and Google agent platforms is planned.

Autonomous cyber resilience roadmap

Cohesity also detailed a longer-term vision called Autonomous Cyber Resilience, which would use agentic workflows to maintain data protection continuously. The company argues that manual triage may not keep pace with attackers using AI to accelerate intrusions.

Under this model, administrators would set goals through Cohesity Copilot, the company's natural-language assistant—such as ensuring critical applications meet specific recovery time and recovery point objectives during an attack. Cohesity Data Cloud would then assess workloads and recommend protection policies, threat scans, and recovery rehearsals, all requiring approval before execution.

One component is available now: customers running Cohesity Data Cloud Enterprise Edition with the company's data security posture management tool can automatically protect newly discovered sensitive data that lacks protection.

Why it matters

As enterprises deploy AI agents with the ability to modify data and systems autonomously, the risk of unintended or malicious actions grows. Cohesity's Global Cyber Resilience Report, released alongside these announcements, found that 55% of IT and security decision-makers feel unprepared to detect or contain unintended agent actions. Agent Resilience represents an early attempt to build recovery capabilities specifically for agent infrastructure, a category that will likely see more attention as agent adoption accelerates.

Cohesity also launched an AI Resilience Academy with a free introductory course on agent risk, available first to Catalyst attendees before joining the broader catalog.

Details were first reported by SiliconANGLE.