Released on October 13, 2026, the CYGNVS Resilience Benchmark Report, developed with Marsh Risk, finds 89% of surveyed organizations have experienced an AI incident, while 46% would improvise their response. The analysis draws on more than 9,500 self-reported surveys and matters because AI now operates workflows with access to documents, email and business systems.
What the survey of 9,500 responses measured
Around 80% of organizations in the dataset use generative AI, yet only one in three test that AI against attacks. Three quarters of AI-using organizations report having an AI governance committee, but the report suggests this does not translate into tested operational defenses. More than six in ten say incident volume increased over the preceding 12 months, while 84% review response plans once a year or less.
The report uses a broad definition of AI incidents that includes harmful hallucinations, bias, model drift, data leakage, agents pursuing unintended objectives, failures between interacting agents and inadequate guardrails. An inaccurate customer-facing answer and an agent exposing confidential information both count, though they are not equivalent security events. The report also says 16% face an AI incident at least weekly, without establishing equal severity across those events.
The methodology covers surveys collected between January 2023 and September 2026 across 21 industries in Asian, European and North American countries. Responses were self-reported, anonymized and analyzed in aggregate, including organizations with and without a critical cyber or AI incident in the prior 24 months. This is a multi-year collection, not a single fresh poll of 9,500 distinct companies, and it does not establish a probability sample representative of all organizations.
What readiness gaps mean for AI operators
Only 33% have a dedicated out-of-band command environment that the AI under investigation cannot observe or influence. If responders keep notes and containment plans in the same documents and messaging an agent can access, the agent may also reach the response. CYGNVS positions its AI Incident Command Center as independent of customer single sign-on, email and infrastructure, with separate identities, Isolate Mode, historical snapshots and signed access certifications.
Only 15% run fully cross-functional tabletop exercises involving legal, executives and communications, while nearly half treat rehearsals mainly as security and IT exercises. Although 88% report having crisis communications plans, more than four in ten have not secured the outside expertise those plans rely on. For companies adopting agents, this points to concrete checks: who can suspend an agent, where to communicate, which records survive, and how specialists join.
The marker to watch is the Resilience Benchmark capability CYGNVS is introducing alongside the report, letting customers compare readiness by industry, size and geography, with a free assessment for noncustomers. If adoption of that benchmarking and retesting grows, boards will get comparable readiness data instead of policy existence alone. CYGNVS sells response-management software, so its commercial interest should be weighed when evaluating the recommendations.
