共有:

IBM Cost of a Data Breach 2026

92% of breached firms had one thing in common:
an access-control gap

Approval workflows and data classification don't help much if SSO, least-privilege, and offboarding are still loose. IBM's latest research found that the overwhelming majority of companies hit by AI-related security breaches had been tripped up by the most basic access controls.

AI Navigate Editorial2026.08.046 min read

POLICY LAYER Approval workflow in place Data classification in place Surface-level governance looks fine ACCESS LAYER 92% lacked basic access control
01
The Finding

What a 604-company survey found:
breached anyway, despite looking "in order"

Approval-workflow and data-classification adoption rates have climbed for years. Breaches still happened.

In August, IBM published its annual research, the Cost of a Data Breach Report 2026: AI Security Edition. Surveying 604 organizations worldwide, the report found that 92% of companies that suffered an AI-related security breach were missing basic access controls — no SSO integration, no least-privilege enforcement, and stale offboarded-employee accounts. Approval workflows and data classification make governance visible, but if the underlying question of who can reach which AI tool, and how far, is left loose, the door stays open for attackers.

Renata Choi, who leads AI risk research at IBM Security, put it bluntly in comments accompanying the report: "Companies have poured so much effort into AI governance paperwork that basic identity and access management keeps getting pushed to later." Among breached companies, 61% had never applied SSO to their AI workloads, and it took an average of 67 days to revoke a departed employee's access to AI tools. The average cost of an AI-related breach rose 8% year over year to $5.54 million, with cases involving an access-control gap running roughly $470,000 higher.

Surface-level governanceEffective access control
Approval workflow in placeSSO + multi-factor authentication required
Data classification policy writtenRole-based least-privilege design
Usage policy and training deliveredImmediate revocation on offboarding
02
Where It Breaks

Access outlives the employee who held it

Permissions granted on day one accumulate with every role change — and stay behind after departure.

Onboarding Access granted via SSO Role change Permissions pile up (least privilege erodes) Departure Avg. 67 days before AI tool access is revoked
FIG. Access accumulates from onboarding through departure, without being reliably reclaimed

What this survey exposes is an amplification effect specific to AI adoption. Agentic AI and RAG stacks are often designed around reaching across multiple SaaS tools and internal databases, so a permission granted once ends up reaching far more data than intended, via the AI tool itself. A suspicious access pattern on a human account tends to get flagged quickly; access via a legitimate service account or API key, on the other hand, tends to fly under the radar — a gap the report calls out directly.

03

The numbers behind the finding

92%
of AI-breached firms lacked access control
604
organizations surveyed worldwide
67 days
avg. time to revoke offboarded access
$5.54M
avg. AI breach cost (+8% YoY)

What got fixed was the visibility — not the foundation.


04
Why It Matters

Why this matters now

Access-management gaps aren't a new theme in IBM's research — it's come up before. What stands out this time is how much more tightly the gap correlates with breaches specifically at AI-adopting companies. The reason: AI tools tend to become a "hub" that reaches across multiple systems and touches large volumes of data. Higher-level governance — approvals, data classification — can look complete while the identity and access layer underneath stays loose, and in that case AI adoption ends up widening the attack surface rather than shrinking it. Read less as a one-off statistic and more as a sign that "AI maturity" is converging with "access-management maturity."

05
Who It Affects

Who it affects, and how

For leadership, the takeaway is that AI governance spend can't stop at policy documents and checklists — it needs to extend to actual investment in IdaaS and access-management tooling. For PMs and IT/systems teams, there's a concrete action available today: add "access control review" as its own explicit line item on the internal AI rollout checklist. For engineers and IT administrators, this is the moment to apply the same least-privilege design and periodic entitlement review to AI agents and API keys that you'd apply to human accounts, and to make offboarding automation a priority.

06
What's Next

Next steps: an access-control checklist

Four items you can start on today.

01

Bring every AI tool under SSO / MFA

Inventory AI tools and extensions still using standalone logins and fold them under SSO. Prioritize unconnected services — the usual breeding ground for shadow AI.

02

Quarterly least-privilege review

Every quarter, audit the permissions that have piled up through role changes and remove what's no longer needed — including service accounts used by AI agents.

03

Automate offboarding

Wire HR offboarding into your identity system so access revocation drops from a 67-day average down to near-immediate.

04

Put it on the AI rollout checklist

Add "access control review" as its own line alongside approval workflow and data classification checks, so it can't be quietly skipped.

07
Risks & Limits

Counter-view, risks, and limits

A few caveats are worth keeping in mind. Much of this survey relies on self-reporting by each company's security team, and the exact bar for "missing basic access control" isn't independently verifiable from public materials alone. The 92% figure also shows a correlation between access-control gaps and breaches, not a proven single cause — other factors, like unpatched systems or weak phishing resistance, are likely compounding in many of these cases. And the 604 surveyed organizations skew toward large and mid-size enterprises, so the picture may not generalize cleanly to small businesses or startups. The best way to put this number to use isn't to over-generalize from it, but to go check your own logs and access-entitlement inventory.