AI Breaches Cost $6M on Average, IBM Report Shows

IBM's 2026 Cost of a Data Breach announcement reported that one in four malicious breaches involved AI-enabled techniques. Those incidents averaged $6 million, compared with $4.99 million across breaches overall. The study examined breaches at 602 organizations between March 2025 and February 2026. The approximately $1 million difference is not a measured governance penalty because the overall average includes AI-enabled incidents and other incident types. The sample cannot establish how much a particular product reduces breach probability, and independent spending evidence is particularly useful because IBM sponsored the study and sells security products.

IBM's 2025 research found that 63% of breached organizations lacked an AI governance policy or were developing one. Among organizations reporting compromised AI models or applications, 97% reported lacking AI access controls, findings that neither describe all companies nor establish a verified 2026 trend. Attacks using AI, such as deepfake deception, may require stronger verification, while attacks against AI assistants may require restrictions on access and execution.

OWASP describes indirect prompt injection as external content changing a model's behavior, illustrated by a fictional supplier document instructing an assistant to export confidential records. Several agents can collectively disclose information even when each agent's individual action appears ordinary. Product evaluation should distinguish governance over purpose and ownership, access limits, data-transfer controls and investigation records, rather than equating incident recording, prevention and recovery. IBM Security Software Vice President Suja Viswesan said AI is making attacks faster and cheaper while breaches become more expensive.

The voluntary NIST AI Risk Management Framework provides a structure for managing AI risks, potentially creating demand for inventories, documented reviews and ongoing evidence collection. IBM's watsonx.governance documentation describes oversight and accountability capabilities, whose commercial value depends partly on operational responses when applications or permissions change. Court Case Finder Web Operations Executive Ben Hartwig recommends testing workflow ownership, information access and stopping authority, including whether changes trigger reviews involving those responsible for access. Continuously updated inventories and tracked exceptions support recurring use, but revenue depends on payment or retention, with separate subscriptions, bundled contracts and consulting producing different economics.

Authentication confirms identity, while authorization determines permitted actions, and OWASP identifies excessive functionality, permissions and autonomy as risks even when credentials are valid. Hartwig recommends checking agents' read and change permissions across connected systems and testing whether access ends when an owner leaves or a task finishes. Okta's September 2026 announcement described available agent identity capabilities alongside planned additions, distinctions that should remain separate from deployment and payment evidence. Contract pricing must connect additional identities, usage or premium capabilities to payment before growth in agent numbers can support a revenue-growth conclusion.

Microsoft's Purview documentation describes sensitive-information controls and investigation of potentially malicious or inadvertent AI-related activity, with support varying by application and configuration. Prevention before a transfer and investigation after disclosure serve different purposes, and evaluation should cover browsers, applications, automated connections, investigation access, retention and alert workload. Hartwig recommends tracing sensitive files through recognition, destination assessment and intervention before transfer, then checking investigation links among users, agents and destinations and testing whether controls disrupt legitimate work.

Gartner forecast in August 2026 that spending on securing AI would reach $4.78 billion in 2027, up 68.7% from approximately $2.84 billion in 2026, rather than representing booked vendor revenue. Gartner Senior Principal Analyst Shailendra Upadhyay said traditional security tools often treat AI applications like other software and need significant updates for AI-specific threats. However, IANS' 2026 research covering more than 500 security leaders found median security budget growth of zero, while 69% identified AI as the leading new-investment priority without distinguishing securing AI from using AI for defense. IANS reported that 38% funded AI outside security budgets, potentially requiring vendors to reach technology or data leaders alongside security teams. An illustrative $100,000 contract replacing $70,000 of software adds only $30,000 to the customer's software budget before implementation costs, showing why replacement spending, bundled renewals and overlapping categories cannot all count as independent new demand.

Okta's revenue rose 11% to $805 million in the quarter ended July 31, 2026. Okta's subscription backlog expected to become revenue within 12 months rose 14% to $2.585 billion. These company-wide figures do not isolate AI-agent revenue, so product-specific contributions, customer expansion and paid adoption are needed to establish the AI connection. Customer evidence should distinguish free trials, production deployments, paid upgrades and existing entitlements, without assuming undisclosed payment arrangements create incremental revenue. Runtime monitoring can require computing, storage and specialist support, making usage pricing, implementation labor, discounts and support costs important to earnings. A small product can grow quickly without materially changing a large supplier's results, so disclosed contributions should be compared with total revenue and cash generation.

An investment watchlist should identify the control gap, product, likely budget owner, payment evidence and financial measure to revisit. Expanding paid deployments, renewals and manageable delivery costs would strengthen the case, while failed trial conversions, unpaid bundling or effective existing tools and procedures would weaken the case. Even strong operating progress may disappoint investors if share prices already assume faster growth.