AI‑powered adversaries and the enterprise risk challenge: Preparing for the new reality | IBM

For the past 21 years, IBM has been working with the Ponemon Institute to gather data from real-life breaches to create the Cost of a Data Breach report. This year’s report studied 602 organizations impacted by data breaches, presenting executives and defenders alike with alarming information that paints the picture of a true tipping point in an AI-powered arms race

Looking at the changes from last year’s report, AI-driven attacks increased by 56%, adding an average of USD 1 million per breach, as attackers use AI tools to increase speed, scale and precision. This is not simply an evolution in attacker tooling; it is a structural shift. When adversaries can automate reconnaissance, generate persuasive phishing content, adapt malware and test exploits at machine speed, the cost and complexity of launching sophisticated attacks drops materially. Breaches become faster, broader and more expensive.

When attack velocity increases, the enterprise has less time to detect, validate and contain an incident. That compressed response window directly drives higher losses, whether through operational disruption, data exposure, legal costs, customer remediation or reputational damage. From the report’s findings, two cost categories, detection and escalation alongside lost business, made up the majority (63%) of costs in the data breaches studied.

Webinar: The rising cost of breaches in the frontier AI era, insights and defense strategies

There is, however, a clear defensive counterweight. AI and automation remain among the most effective ways to reduce breach impact, helping security teams move at machine speed and delivering an average savings of USD 1.93 million per breach. That number deserves attention in the C-suite. It indicates that AI is not simply another discretionary technology investment within the security stack, but rather a core control mechanism for limiting business impact and financial exposure.

This year’s report also looked at how organizations treat this need for investment and direction. We have found that adoption remains uneven across the security lifecycle. While 50% of breached organizations have deployed AI agents in threat hunting, response and containment, only 18% have applied them to vulnerability scanning and management. That gap is significant. It suggests many enterprises are still using AI reactively, after suspicious activity has emerged, rather than proactively, where frontier capabilities deliver outsized advantages.

Another important finding is that roughly one in five organizations reported an AI-related breach—that is, an attack on AI models or applications. Of those, the vast majority (92%) lacked proper AI access controls, and the underlying causes are telling

The incidents that involved AI were less about model selection and more about model and environmental security. The most expensive types of security incidents involving an organization’s AI model or applications were inversion and prompt injection attacks, which led to average losses of USD 6.07 million and USD 5.89 million, respectively. A model inversion attack occurs when adversaries exploit an AI model to infer or reconstruct sensitive training data, such as personal, proprietary or confidential information, by analyzing model outputs and responses. These attacks can expose underlying data without direct access to the original dataset.

In terms of environment types, similar security incident rates were observed across both open-source and third-party vendor deployments, which suggests the core problem is not simply whether organizations build or buy their AI capabilities, but how they manage and secure them once they are deployed. This was apparent from root causes that frequently included familiar issues: compromised connected APIs, vulnerable applications and cloud misconfigurations. In other words, many AI-related breaches stemmed from structural weaknesses in the enterprise environment rather than flaws inherent to a specific model.

This distinction matters for senior leaders. It reframes AI security from a narrow model-risk conversation into a broader operating-model challenge involving architecture, controls, accountability and oversight

Encouragingly, organizations appear to recognize this. More than half now say they plan to invest in AI security and governance tools post-breach, representing an 88% increase from last year. That shift reflects growing awareness that securing AI requires securing the entire ecosystem around it

In response to emerging threats from frontier AI models, 85% of breached organizations say they plan to increase spending on security tools and governance. That is a rational and necessary response, but increased spending alone will not be sufficient unless it addresses one of the fastest-growing risk areas: non-human identities (NHIs)

As AI agents proliferate across business and security workflows, they create new identities that interact with applications, APIs, data stores, orchestration tools and cloud services. These non-human identities often operate with elevated privileges and can be difficult to inventory, monitor and govern. Yet this year’s findings show that fewer than half of organizations are actively securing non-human identities.

For CISOs and those managing AI use across the enterprise, this is one of the immediate priorities. AI workflows are only as secure as the identities, permissions and execution environments that support them. Without strong controls over machine identities, secrets management, access boundaries and auditability, organizations risk introducing a new class of hidden attack paths into already complex ecosystems.

Rising through the report’s findings is the criticality of the AI sovereignty concept, which entails maintaining control over where AI systems run, how data is processed and who has access. For global enterprises operating across jurisdictions, this is becoming a crucial security and governance requirement. But, as AI systems scale and embed deeper into organizational workflows, identity and access controls have failed to keep pace. In this year’s report, only 40% of organizations reported using access controls on AI models and data.

Without AI sovereignty, organizations expand their attack surface and create dependencies that are difficult to monitor or control. Sensitive data may traverse environments with inconsistent protections. Model operations may depend on external platforms without adequate visibility. Regulatory exposure can increase as operational resilience decreases

For enterprise leaders, AI sovereignty should not be viewed only through a compliance lens but rather as a cyber resilience issue. Control over infrastructure, data handling, identity and access is foundational to reducing both concentration risk and systemic exposure, thereby enabling teams to better control the outcomes of data breaches. This is where executive sponsorship becomes critical as teams across various parts of the organization must join hands in managing a rapidly evolving strategic AI battleground.

For the 2026 Cost of a Data Breach report, the core message for the C-suite is simple: AI is compressing the time between exposure and impact. In that environment, cyber leadership must move from incremental modernization to strategic acceleration. The tipping point is not coming—it has arrived

For more insights, advice and recommendations, read the full Cost of a Data Breach report here. Get a breakdown of the report from industry experts in the webinar, The rising cost of breaches in the frontier AI era, insights and defense strategies—register here

Cost of a Data Breach report 2026

The global average cost of a data breach reached USD 4.99M while AI-driven attacks increased 56%. Explore the latest findings

Source: www.ibm.com