A survey of more than 9,500 organizations in 21 industries across Asia, Europe and North America, collected from January 2023 through September 2026 by CYGNVS and Marsh, found that 89% have already faced some kind of AI incident. Sixteen percent face one at least weekly. The incidents include rogue agents, agentic runaway, data leakage and broken guardrails.
About 80% of organizations use generative AI. Three out of four of those have an AI governance committee, but only one in three test their AI against attacks. Organizations use AI for marketing nearly three times as often as for risk assessment, so the fastest adoption is in the functions least likely to be tested or governed.
Company size does not close the gap. Only 39% of companies with $1B or more in revenue test their AI adversarially. In the $50M to $1B revenue bands, the figure is 29% to 31%.
The response plan has a blind spot
When an AI incident does occur, 46% of organizations would improvise their response. Only 33% have a dedicated out-of-band command environment that the AI under investigation cannot observe or influence. Another 21% can run a response somewhat separately, but their logging and isolation are not purpose-built. The remaining organizations would improvise isolation or fall back on the same email and chat systems the AI may have access to.
Ownership is also unclear. 67% of organizations have not mapped clear owners for their most critical processes and dependencies, and 53% would still be assigning roles and access after an incident starts.
Practice and reporting trail the threat
Incidents are rising. More than 6 in 10 organizations say volume increased over the past 12 months, and 91% of CISOs said their board became more involved. Yet 84% review their incident response plans once a year or less, 17% do not run tabletop exercises at all, and 66% run them only every one to two years. 67% have never tested recovery under real incident conditions.
Reporting obligations are growing faster than the processes to meet them. More than 50 jurisdictions now have laws or rules on AI incidents, but 52% of organizations have no defined process for the notifications those rules require. 90% are now required to formally document security incidents, up from 82% in 2023, and 69% admit significant gaps in their records of who did what and when.
Outside help is another weak point. 88% have a crisis communications plan, but more than four in ten have not identified the PR, forensics and outside counsel providers that plan depends on. 18% still share incident materials with outside experts by plain email.
What organizations that have been through an incident do differently
Organizations with a recent critical incident were further along at almost every stage. 63% of them use a true single source of truth during a live incident, compared with 48% of those without one. 71% report on cyber resilience to the C-suite and board at least monthly, compared with 38%. Only 31% of them run security-only tabletop exercises, compared with 49%.
The report’s recommendations for organizations that want those results without the incident:
- Treat AI incidents as their own category, with expert playbooks for rogue agents, agentic runaway, bias, data leakage and missing guardrails, and adversarial testing of AI systems.
- Run tabletop exercises more often, with legal, risk, compliance, communications and executives as full participants, and test recovery under real conditions.
- Run the response from one secure, out-of-band environment, and lock in outside experts before the next incident starts.
- Build a record of who knew what and when, map disclosure obligations across every applicable jurisdiction, and put a process in place for AI incident notifications before a regulator asks.
The full CYGNVS Resilience Benchmark Report is at CYGNVS.com/resilience.
