Cohesity launches Agent Resilience for AI recovery

Cohesity has launched Agent Resilience, a new feature in its Data Cloud platform designed to protect and recover AI agent infrastructure. The launch follows company research showing that 56% of organisations do not feel well prepared to recover from unintended actions by AI agents.

The product is designed to protect the systems behind enterprise AI agents, including agent memory and configuration, and help organisations restore them to a known good state after corruption, misconfiguration or malicious activity. It also aims to safeguard the databases, file systems and other services those agents can access and alter.

The launch reflects growing concern that businesses are giving AI agents access to operational systems and privileged credentials without equivalent recovery plans if something goes wrong. Many organisations also still lack a clear inventory of AI agents and workflows running across their operations, making recovery planning harder.

Research cited by Cohesity found that 58% of organisations were not very confident in their ability to verify the integrity of AI models and related data after a cyberattack. In Australia, 47% said they did not have a centralised inventory or clear understanding of the AI agents, copilots and workflows operating across their business.

At launch, Agent Resilience supports Amazon Bedrock AgentCore and Amazon Bedrock Agents. Support for Microsoft and Google agent platforms is planned.

The product uses snapshots, immutable backups and clean room recovery, which Cohesity already applies to other workloads across on-premises, cloud and software-as-a-service environments. It also offers what it calls agent topology, giving teams a single view of each agent, its memory stores, connected applications, databases and related infrastructure.

That visibility is intended to help organisations understand dependencies and determine what must be restored to resume trusted operations after an incident. The service is available now to select customers, with wider availability targeted by the end of the year.

Vasu Murthy, Chief Product Officer at Cohesity, said recovery remains a weak point in many AI oversight programmes.

"Detection can tell you an AI agent went off course. It cannot undo the changes," said Vasu Murthy, Chief Product Officer at Cohesity.

"Cohesity Agent Resilience provides a precise recovery path for both the agents and the resources they impact. Our vision for Autonomous Cyber Resilience is to use agentic workflows to reduce manual work in response and recovery," Murthy said.

Broader strategy

Alongside the launch, Cohesity outlined a broader strategy it calls Autonomous Cyber Resilience. The model would use agentic workflows to automate parts of the company's five-step cyber resilience framework, covering protection, recoverability, remediation, recovery rehearsal and risk posture improvement.

The aim is to move organisations away from manually configured plans reviewed only at intervals. Instead, Cohesity wants systems to continuously discover new assets and data, assess whether they are protected, recommend policy changes and support incident response, with a human still approving key actions.

In practical terms, administrators would define business recovery objectives through Cohesity Copilot rather than setting policies application by application. The platform would then assess resilience posture, suggest protection and scanning approaches, and recommend recovery rehearsal plans for approval.

During an incident, those workflows could help identify the scope of compromise, detect signs of attacker activity and prepare an isolated recovery environment. Cohesity argues this matters because attacks are becoming faster and more automated, while many recovery processes still rely on manual triage and sequential approvals.

Customers with Cohesity Data Cloud Enterprise Edition and its data security posture management product can now also automate protection for newly discovered sensitive data that is not already covered. The process includes continuous discovery and classification, checks on protection status and policy-based trigger actions.

Partner view

Cognizant, which works with enterprises on technology deployments, said the ability to recover AI agent state could become more important as those systems move into operational use.

"As AI agents take on more business-critical work, we need to think beyond monitoring and make sure we can recover when something goes wrong," said Srikanth Kuntamukkala, CIS Global Lead for Partner Sales at Cognizant.

"Being able to restore an agent's state, and the data it may have changed, could give our teams a more practical way to manage that risk as we move agentic applications into production," Kuntamukkala said.

Training push

Cohesity also introduced AI Resilience Academy, a training track within its existing academy. Its first course focuses on the distinction between AI assistants, which make recommendations, and AI agents, which can take action.

The free, self-paced course is intended to help organisations understand the risks that arise when AI systems move from advisory roles to making changes in live environments. Topics include defining AI agents, identifying gaps in the AI lifecycle and understanding upstream and downstream resilience issues.

Murthy said the broader challenge is not limited to a single product category.

"Cyber resilience is not a one-time assessment. It requires teams to protect data, test recovery, detect threats, rehearse responses, and continually improve their risk posture," Murthy said.

"Agentic workflows can take on more of that routine work and help keep the plan current," he said.