AI-Driven Breaches Cost US$1M More. Can You Contain Them Fast Enough?

AI is helping attackers move faster and making breaches an average of US$1 million more expensive.

The IBM Cost of a Data Breach Report 2026 found that the global average reached a record US$4.99 million, up 12% in a year. In Australia, the average breach has reached US$2.96 million.

The report's other numbers make for uncomfortable reading:

• More than one in four malicious breaches were AI-driven, a 56% increase from the previous year.

• Deepfake impersonation accounted for 45% of AI-driven attacks, followed by AI-enabled malware at 19% and AI-generated phishing or other communications at 17%.

• Security teams using AI and automation extensively saved US$1.93 million on average and identified and contained breaches 65 days faster than organisations using none.

• Shadow AI was involved in 43% of security incidents affecting an organisation's AI environment.

Attacks have accelerated while many investigations still depend on fragmented tools, manual triage and a queue.

The Bill Rises When Response Slows

Detection and escalation, together with lost business, accounted for 63% of the global average breach cost. These are the costs that grow when a threat is difficult to find, takes too long to understand or interrupts the organisation while the response team works out what happened.

The average breach took 247 days to identify and contain, up from 241 days the year before. Breaches lasting longer than 200 days cost US$5.65 million on average, compared with US$4.32 million for shorter incidents.

An alert can arrive in seconds. Understanding whether it matters takes context: which identity was involved, what system was touched, what data was available and whether the behaviour belongs there. When those answers sit across separate tools and teams, investigation time becomes breach cost.

The report also gives people their due. Internal IT and security teams identified and contained breaches in 209 days on average. Managed security service providers did so in 230 days. Both were faster than the global average. Automation helps, but someone still has to make the call.

AI Risk Sits Around the Model Too

More than 20% of organisations in IBM's study reported a breach targeting AI models or applications. Much of the exposure sat in the surrounding systems: compromised application programming interfaces, applications or plug-ins accounted for 27%, as did cloud misconfigurations affecting AI workloads.

An approved AI platform can still expose information through everything connected to it. Employees use unapproved tools. Developers connect assistants to source code and credentials. Custom applications and agents receive access to business systems. Each connection adds another identity, permission and data path to account for.

A useful AI security picture answers five questions:

• Which tools, assistants, applications and agents are in use?

• What information can they access or send elsewhere?

• Which identities and permissions sit behind them?

• Can unsafe prompts, responses or automated actions be identified and controlled?

• Can an incident be traced across identity, endpoint, network, cloud and data systems?

Shadow AI is already part of the environment. Treating it as a future governance problem leaves today's security team protecting systems it may not know exist.

Faster Defence Needs Connected Operations

Aussie organisations are adopting AI across everyday work, which calls for connected visibility, governance and response. The operating model is fairly plain.

AI Use Is Visible and Governed

Mature organisations maintain a current view of approved and unapproved AI tools, developer assistants, custom applications, agents and Model Context Protocol (MCP) servers. Policies reflect the user, tool, data and action, rather than relying on a static list of permitted platforms.

Data and Access Are Kept Narrow

Sensitive information is classified, unnecessary access is removed, and human and non-human identities receive only the permissions they need. Prompt, response and agent activity can be reviewed when something falls outside policy.

Detection Leads Somewhere

Identity, endpoint, network, cloud and application signals feed an operating model where analysts can connect the evidence and act. The useful measure is the time from a meaningful signal to a defensible decision.

Recovery Is Planned Before It Is Needed

Only 42% of organisations in IBM's study had fully recovered from their breach. Recovery means more than containing an attacker. Operations need to return to normal, compliance obligations need to be met, and customer and employee trust needs to be rebuilt. That work is easier when ownership and decision paths are settled before an incident.

Defenders Need the Same Speed

IBM's report makes the advantage clear: prepared defenders contain breaches sooner and at lower cost.

An AI-enabled breach is not a single event but a chain. A malicious file or process lands on one endpoint, credentials or data are exposed, and the activity moves across servers or cloud workloads before the original signal is connected to the wider incident.

Lumara Sentry picks up the endpoint signals in that chain across laptops, servers and cloud workloads. It works within Lumara SecOps Cloud and is supported by Lumara Operate, our 24/7 Australian SOC, which validates findings, cuts through the noise and helps stop malicious processes, isolate affected devices and contain threats before they spread.

AI-driven attackers are compressing the time between first access and impact. If one reached an endpoint in your environment today, how quickly could you detect and contain it? Stop endpoint threats before they spread with Lumara Sentry.