Most AI programs begin with the model. The breach data suggests they should begin with access. IBM's 2026 study found that one in four malicious breaches were AI-enabled, a 56% increase over the prior year, costing an average of USD 6 million.1 IBM's first recommended action is to secure and control agentic identities.2
The warning signs were visible a year earlier: in IBM's 2025 report, 97% of organizations that suffered an AI-related breach lacked proper AI access controls.3
Why agents break old assumptions
- They act with delegated authority. An agent booking travel or closing tickets uses someone's permissions. Delegation must be explicit, scoped and revocable.
- They scale instantly. Gartner projects more than 150,000 agents in an average Fortune 500 enterprise by 2028.5 Manual access reviews cannot keep up.
- They chain together. One agent calling another creates access paths no one designed.
An identity-first blueprint
Give every agent a distinct identity, never a shared service account. Issue short-lived, narrowly scoped tokens through standard OAuth flows. Bind each agent to an accountable human owner. Log every action against that identity so audits can answer who did what, and on whose behalf.
Microsoft finds 88% of enterprises already experimenting with agents.4 The organizations that put identity controls in place now will scale agents with confidence. The rest will discover their access model in an incident report.
Related insights
Friction, fraud and the customer at the login screen
Every extra step at sign-in costs conversions. Every missing step invites fraud. The data on where to draw the line.
Twenty-nine minutes: how fast attackers move once they are in
Breakout time keeps shrinking. Identity telemetry is the fastest signal defenders have.
Identity is the new perimeter. The 2026 breach data proves it.
Four of the year's most respected threat reports point the same way: attackers would rather sign in than break in. Here is what that means for every security budget.