Cohesity survey finds 39% of companies have recovery plans that fully cover attacks on AI

Rolling back the systems an agent damaged does not roll back the agent ^[9]. Long-running agents accumulate memory, configurations and database values ^[8]. When bad code, data poisoning or another fault changes how one behaves, the agent itself has to go back to a point when it worked correctly ^[8]. Cohesity's report, as SC World describes it, says recovery has to run both ways. Business systems go back to the state before the agent's erroneous actions, and the agent goes back to a known-good state ^[9].

Rob Sadowski, Cohesity's vice president of product and solution marketing, put it as questions in an interview with Enterprise Security Weekly's Jackie McGuire ^[12]. "What does the agent have access to? And if it makes a mistake or makes an error, does some unwanted action, can I go and roll those systems back to a state before the agent made that error?" he said ^[12]. "Can I actually roll back the agent itself back to a point in time when it was functioning according to spec?" he said ^[13].

For a responder, the second question decides when the incident is over. Restore the database but leave the agent's memory untouched, and whatever drove the bad writes is still there when the agent resumes ^[8]. No intruder is required. The scenario the report describes is an agent doing something destructive without being compromised ^[3].

The gap between AI use and full plan coverage is 60 percentage points ^[18]. That leaves about 1,950 of the 3,200 respondents who did not say their plans comprehensively accounted for attacks on AI systems, workflows or models ^[17]. Sadowski's summary measured something slightly different: formal inclusion of AI systems in recovery plans, where the survey line asks whether plans account for attacks on them ^[10][2]. "Less than half of organizations that we talked to had formally included their AI systems into their recovery plans," he told McGuire ^[10]. "We have more and more of our critical operations being influenced or even driven by AI, yet they don't have it incorporated in the plan," he said ^[11].

The piece also argues that attackers use AI to automate operations and find vulnerabilities, cutting the time defenders have to respond ^[14]. It adds that agents and MCP servers bring more connections and dependencies to an attack surface that was already complicated ^[15]. Every figure here is self-reported and comes from Cohesity's own report, with a Cohesity marketing executive presenting the findings ^[1][10]. The article does not describe how respondents were sampled, does not define "comprehensively", and does not cite an incident where missing AI coverage made a recovery take longer.

The rollback argument stands without the survey. An agent that keeps state can carry a fault through a clean restore of everything around it ^[8]. The urgency, by contrast, rests on a vendor's self-reported numbers ^[1].

Cohesity recommends starting with a "Minimum Viable Company": the smallest combination of applications, data and dependencies needed to keep critical operations running. AI systems and their dependencies get mapped into that set, and the plan gets tested repeatedly ^[16].