Key Takeaways

  • IBM found that one in four malicious breaches involved AI, while AI-driven attacks increased 56% year over year.
  • AI-enabled breaches averaged $6 million, compared with a $4.99 million global average across breaches.
  • Organizations using AI-enabled security reported reducing the financial impact of an incident by an average of $1.93 million (source).

Artificial intelligence has moved deeper into the economics of cybercrime. New 2026 findings from IBM and the Ponemon Institute show that AI-enabled data breaches cost organizations an average of $6 million, roughly 20% above the $4.99 million global average across breaches.

The research examined 602 organizations worldwide between March 2025 and February 2026. One in four malicious breaches involved AI, and the volume of AI-driven attacks increased 56% year over year. Those figures indicate the technology has moved rapidly from experimental use to routine offensive operations.

Attackers can use generative AI to produce convincing phishing messages, impersonate executives and suppliers, research potential victims, discover vulnerabilities, and accelerate parts of malware development. None of those tactics is entirely new. What changes is the speed, volume, and level of personalization available to a criminal group.

That compression of time matters. A phishing campaign that once required substantial manual work can be adapted across languages, job roles, and industries much faster. Fraudulent messages can also draw on publicly available information to sound more credible. How many employees will pause to investigate when a request looks familiar, arrives in polished language, and appears to come from a senior executive?

AI is not only an attack engine, but also an expanding attack surface.

Enterprise systems such as Microsoft Copilot and Google Gemini can interact with corporate documents, email, internal knowledge bases, and other business applications. If access controls are too broad, an AI assistant may retrieve information that a particular user should not see. Prompt injection can attempt to manipulate a model's behavior, while poorly governed integrations can create paths for data leakage.

Permissions are therefore becoming a central part of AI security. Traditional identity controls still apply, but the consequences of excessive access can become harder to predict when an AI system searches across multiple repositories and summarizes the results. A user might not know where sensitive information is stored. The assistant may find it anyway.

The gap between concern and defensive adoption remains significant. BCG reported in 2025 that about 60% of senior leaders believed their organizations had likely experienced an AI-powered attack, yet only 7% were using AI-enabled defenses. That imbalance gives attackers room to automate while many security teams remain dependent on manual triage and fragmented tools.

IBM's results also offer evidence that defensive AI can improve the financial picture. Organizations using AI-enabled security reported reducing breach impact by an average of $1.93 million per incident. Vendors including CrowdStrike and Palo Alto Networks are applying AI to areas such as behavior analysis, alert prioritization, threat detection, and incident investigation. Human oversight still counts, of course. An automated system that acts on weak data or excessive privileges can generate a different category of risk.

Asset discovery is a practical starting point for businesses evaluating their defenses. Security and technology leaders need visibility into approved AI products, embedded AI functions within existing applications, employee use of public models, connected data repositories, and the identities authorized to access them. Shadow AI complicates that inventory because employees can adopt services before security teams have evaluated their data handling.

Governance frameworks can help translate that inventory into controls. The NIST draft Cyber AI Profile, identified as NIST IR 8596 in 2025, organizes the challenge around securing AI systems, using AI for cyber defense, and thwarting AI-enabled attacks. It complements NIST Cybersecurity Framework 2.0 rather than treating AI security as an isolated discipline.

In practice, companies can apply least-privilege access, monitor prompts and outputs where appropriate, test models for prompt injection, restrict sensitive data flows, and establish incident procedures for AI-related events. Procurement teams also have a role. Questions about model hosting, retention, training data, audit logs, third-party integrations, and breach notification belong in vendor reviews.

The broader message from IBM is uncomfortable but useful. AI is increasing breach costs while helping defenders reduce them. The advantage may depend less on who adopts AI first and more on who governs access, detects misuse, and responds at machine speed without surrendering human judgment.