IBM’s 2026 breach report shows where data security breaks down

At a glance

What this is: IBM's 2026 breach report argues that modern data losses are happening in the seams between tools, especially where AI workloads, shadow AI, and partial encryption coverage create ungoverned gaps.

Why it matters: This matters because identity, data, and AI controls are increasingly interdependent, so IAM, PAM, and NHI teams need to understand where coverage stops and where enforcement must shift.

By the numbers:

• IBM's 2026 Cost of a Data Breach Report put the global average breach cost at $4.99 million, up 12% from a year earlier.
• The share of security incidents involving shadow AI rose to 43% year over year.
• Only 37% of breached organisations encrypt sensitive data both at rest and in transit.

Context

Data security fails when controls cover isolated states or channels but do not cover the full path a record takes across endpoints, cloud services, SaaS apps and AI tools. In this report, the central issue is not a lack of security products, but fragmented ownership and visibility across the places where data actually moves.

The identity angle is increasingly important because data access is now mediated by human users, applications and AI agents operating with delegated permissions. When access control, encryption, data discovery and AI governance are treated as separate programmes, gaps emerge exactly where attackers and unsafe usage can pass through unnoticed.

Key questions

Q: What breaks when data security tools only cover part of the workflow?

A: Coverage breaks at the handoff points. A tool can protect data at rest, in transit or on the endpoint, but if the record moves into an unmanaged AI tool or another unmonitored channel, the organisation loses continuity. The failure mode is not one bad control, but an unowned transition between controls.

Q: Why do shadow AI tools create risk even when endpoint and DLP controls exist?

A: Shadow AI creates risk because those controls depend on sanctioned channels and known destinations. If a user pastes sensitive content into an unsanctioned chatbot, the organisation may never see the transfer, the retention terms or the downstream reuse. The problem is governance loss, not just visibility loss.

Q: How can organisations tell whether AI governance is actually working?

A: Organisations can tell AI governance is working when they can inventory every agent, explain its purpose, show who owns it, and prove that permissions are tightly scoped. If those four things are missing, the programme has policy language but not operational control. Auditors will notice the gap quickly.

Q: What should organisations do when AI tools can take actions on behalf of users?

A: Treat agent actions as governed identity events and require logging of tool calls, delegation chains, and execution context. Without that layer, investigations cannot tell whether an action came from the human operator, the agent, or a downstream delegated workflow.

Technical breakdown

Why point tools miss AI-era data movement

Point tools usually specialise in one control plane, such as data at rest, data in motion, endpoint activity or cloud posture. That model breaks down when the protected object is not a static file but a record moving through SaaS, browser sessions, chat interfaces and AI workloads. A control can be technically correct and still fail operationally if it only sees one slice of the workflow. In AI-heavy environments, the data path is now dynamic, user-mediated and often transient, which means the security question is coverage continuity rather than isolated product capability.

Practical implication: Map where each data control begins and ends, then identify the unowned handoffs between endpoint, SaaS, browser and AI workflows.

How shadow AI creates governance gaps

Shadow AI is not just unsanctioned software use, it is an access and data-governance problem. When users paste sensitive information into unmanaged AI tools, the organisation loses visibility into the destination, retention behaviour and secondary use of that information. Traditional DLP may miss the event if the channel is outside sanctioned traffic, and data-at-rest controls cannot enforce policy once the record leaves governed storage. The governance gap widens when no inventory exists for the tools and agents that are actually handling data.

Practical implication: Inventory approved and unapproved AI tools, then enforce data controls at the point of interaction instead of relying on storage-side inspection.

Identity and data security are converging around AI agents

An AI agent with delegated permissions can touch data in ways that resemble a human user, but at machine speed and with less predictable behaviour. That makes identity governance part of data security, because access control is no longer only about people and services, but also about software entities that can act on behalf of people. If a control stack cannot distinguish a person from an agent, or cannot scope what each is allowed to read, move or disclose, it will miss a major part of the risk surface.

Practical implication: Treat AI agents as governed identities and review whether their data permissions are scoped, monitored and revocable like other privileged actors.

Threat narrative

Attacker objective: The objective is to obtain sensitive business data through ungoverned AI or workflow paths that are outside the organisation's monitored controls.

1. Entry occurs when a user, application or plug-in routes sensitive data into an unmanaged AI tool or poorly governed workload.
2. Credential or permission abuse follows when the workload or agent can access data beyond the original user intent through delegated access or weak governance.
3. Impact occurs when the record leaves monitored channels and the organisation loses visibility into where it was stored, copied or reused.

Breaches seen in the wild

• Hugging Face Spaces breach 2024: Unauthorised access to Hugging Face Spaces may have exposed secrets users stored for AI apps; tokens were revoked and org tokens removed.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Incomplete coverage is now the dominant data security failure mode. The report's most important signal is not that a single control failed, but that the control stack stopped at the seams between data states, AI workloads and delegated access. That means the governance problem is architectural, not cosmetic. Practitioners should stop asking which tool failed and start asking which handoff is unowned.

Shadow AI is an identity governance problem as much as a data problem. Once users interact with unmanaged AI tools, the organisation no longer controls who or what is processing the data, how long it persists, or whether it is being reused elsewhere. That makes AI inventory, access scoping and policy enforcement part of the same risk conversation. Practitioners need governance over the actor, not just the data object.

AI agents widen the gap between permission and behaviour. A delegated agent can exercise access in ways that look legitimate at issuance time but become difficult to reason about after the fact. This creates a governance trust gap where classic review processes may be too slow or too static. Teams should treat agent permissions as runtime risk, not just entitlement records.

Encryption coverage is only meaningful when it spans the full data path. Encrypting at rest without comparable protection in transit leaves the transition point exposed, and that transition is where many modern workflows now live. The report reinforces a broader lesson for identity and data programmes: coverage that depends on separate owners and separate tools will always leave operational seams. Practitioners should measure control continuity, not just control presence.

What this signals

Data security programmes need to move from tool coverage to control continuity. The practical question is no longer whether a stack includes DLP, DSPM or endpoint protection, but whether those controls remain effective when a record crosses from one environment to another. Teams should measure the seams, because seams are where governance fails first.

AI governance is becoming a data governance dependency. Once employees and agents can move information through unsanctioned or semi-sanctioned AI services, identity policy, data policy and acceptable-use policy stop being separate topics. Practitioners should align those programmes before the next tool rollout creates another blind spot.

Policy must be enforced where interaction happens, not only where data is stored. That means browser, endpoint and SaaS touchpoints deserve the same scrutiny as repositories and vaults. In environments with AI agents, this also means defining who is allowed to act, what they can touch and how those permissions are withdrawn.

For practitioners

• Map control boundaries across the full data path Document where your DLP, DSPM, endpoint, SaaS and AI controls each begin and end, then flag every handoff that depends on a different team or product.
• Inventory shadow AI and unmanaged assistants Build an inventory of approved and unapproved AI tools, browser-based assistants and plug-ins that can touch sensitive data, then classify them by business use and data exposure.
• Scope AI agent permissions like privileged access Review AI agents and automations as governed identities, with explicit data-read, data-move and data-disclose permissions that can be revoked when the use case changes.
• Test encryption continuity across states Verify that sensitive records remain protected not only in storage but also during transfer between email, browser sessions, SaaS apps and AI services.

Key takeaways

• The report's core lesson is that modern breaches exploit coverage gaps between tools, not just weaknesses inside a single control.
• AI-enabled malicious breaches averaged $6 million, and shadow AI incidents rose to 43%, showing that unmanaged AI usage has become a measurable security exposure.
• Teams need continuous control across data states, AI workloads and delegated identities, or they will keep defending the wrong boundary.

Key terms

• Shadow AI: AI agents, copilots, or connected tools operating without full visibility or governance from security teams. Shadow AI becomes an identity problem when those systems authenticate with unmanaged tokens, service accounts, or OAuth apps that can reach production resources.
• Control Seam: The transition point between two controls, teams, or processes where ownership can become unclear. These seams are often where incidents progress, because prevention, detection, and response each assume another function will catch the issue.
• Delegated Identity: Delegated identity is when one actor acts on behalf of another with explicit permission and bounded authority. In AI-assisted commerce, it requires clear consent, limited scope, and traceable records so the retailer can distinguish authorised delegation from unauthorised automation.
• Data Security Continuity: Data security continuity is the ability to protect information consistently as it moves across storage, endpoints, cloud services, email and AI tools. It matters because many breaches exploit transitions between states, not weaknesses in a single storage location.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, agentic AI identity and secrets management. It helps practitioners connect identity control with the wider security decisions their programmes have to make.