Skip to main content
Process diagram: how an approved AI service drifts away from the access model it was approved under. Illustration: SecurityXP

AI Approval Is Not Access Control: What the 97% Breach Statistic Actually Says

· 5 min read · SecurityXP Editorial Desk

Key Facts

  • IBM's 2025 Cost of a Data Breach Report studied 600 organizations breached between March 2024 and February 2025.
  • 13% of studied organizations reported a breach involving AI models or applications; 97% of those said they lacked proper AI access controls.
  • 60% of AI-related incidents led to compromised data and 31% led to operational disruption.
  • 63% of breached organizations lacked an AI governance policy or were still developing one, and 20% reported a breach involving shadow AI.
  • Claude's enterprise controls include single sign-on, SCIM provisioning, routing rules, roles, group mappings and admin API keys.
  • Rapid7's Surface Command integration imports Claude organization, user, role, group and settings data through the Compliance API, but does not ingest conversation content.

IBM’s 2025 Cost of a Data Breach Report, conducted by Ponemon Institute, studied 600 organizations that suffered a breach between March 2024 and February 2025. Thirteen percent of them said the incident involved AI models or applications. Among that group, 97% said they lacked proper AI access controls.

The figure is striking, but it is also easy to misread. It does not mean 97% of all organizations are exposed, or that access control alone would have stopped every incident. It means that almost every organization in the study that reported an AI-related breach had not put basic identity and access controls around the AI systems involved. That is a narrower claim, and a more useful one.

What the report measured

The report covers 600 breached organizations across 17 industries and 16 countries. The AI findings are a first for the series: previous editions tracked AI as an attack tool, not as a system with its own access model.

The numbers behind the headline matter:

  • 13% reported a breach involving an AI model or application; another 8% were unsure whether AI was involved.
  • Among those reporting an AI-related breach, 97% said they lacked proper AI access controls.
  • 60% of the AI-related incidents led to compromised data; 31% caused operational disruption.
  • 63% of breached organizations either had no AI governance policy or were still developing one.
  • 20% reported a breach involving shadow AI, the use of AI tools without employer approval or oversight.

The sample is self-reported and not statistically representative, and AI incidents remain a small share of the total. But the direction is consistent with what security teams see in practice: AI services are adopted faster than the identity controls around them are updated.

Approval is a procurement decision

Approving an AI service is usually a decision about whether the organization may use it. It may cover contract terms, data handling, a security questionnaire and a named business owner. It rarely answers the questions that determine who can actually reach the system six months later.

Those questions live in the identity layer. In Anthropic’s enterprise Claude documentation, the control plane includes verified domains, single sign-on through SAML or OIDC, SCIM provisioning, routing rules, directory groups, seat tiers, roles and admin API keys. Each is a place where the approved state and the current state can diverge.

Consider a few concrete examples from that documentation:

  • Without SCIM, an account is created the first time someone signs in. Deprovisioning then depends on a manual process rather than the directory.
  • Group mappings assign seat tiers and roles. Reordering groups can change a person’s tier without waiting for the next directory sync.
  • Moving a user between organizations ends their active sessions, but it does not revoke admin API keys they created.
  • If a directory-managed account is not yet linked, a deactivation request is refused until the account is linked. The person can remain reachable in the meantime.

None of these is a flaw in Claude specifically. They are the normal mechanics of enterprise identity, and they are exactly why an approval record is not an access review.

Where access drifts

Access drift is rarely a single dramatic event. It accumulates through ordinary changes:

  • A person changes role, but their group membership still reflects the old job.
  • A contractor leaves, but an API key or service credential remains active.
  • A team adopts a new AI connector, and the AI service can now reach data the approver never considered.
  • An employee uses an unapproved AI tool, creating a shadow account outside the identity provider entirely.
  • An admin leaves the company, and the break-glass path is the only remaining owner.

The IBM data supports the scale of the problem, not the specifics of any one organization. The 97% figure is a warning that AI systems are often outside the access-review cadence that covers email, VPN and finance applications.

Six questions for an access review

A useful review starts with the identity provider, not the AI vendor’s user list. Six questions cover most of the gap:

  1. Who has an account, and how was it created? Separate directory-managed accounts from accounts created on first sign-in.
  2. Who holds owner or administrative roles? List every Primary Owner, tenant administrator and organization administrator, and confirm each still needs the role.
  3. Which groups map to which roles and seat tiers? Check the mapping rules, not just group membership. Reordering or renaming a group can change access.
  4. Which credentials are not tied to a person? Admin API keys, service accounts and integration tokens need an owner, a purpose and an expiry.
  5. What happens when someone leaves? Test the leaver path with a real account, including API keys and connected applications.
  6. What can the AI service reach? Review connectors, plugins and data sources. Access to the AI service is not the same as access to the data behind it.

The answers should be reviewed on a schedule, not only when a service is approved. A quarterly review is more useful than a one-time questionnaire.

Visibility tools help, but they do not decide

Vendors are starting to fill the visibility gap. Rapid7’s Surface Command, for example, now imports supported Claude organization, user, role, group and settings data through Claude’s Compliance API. That gives security teams a way to inspect AI identities alongside other assets.

The limits are explicit in Rapid7’s own description: the integration does not ingest conversation content, provide threat detection on Claude activity, or control Claude’s behavior. It supports a review; it does not perform one. It is also a vendor integration, not an independent audit.

That distinction matters. Buying a visibility product without assigning an owner, a review cadence and a remediation path leaves the same gap in place, only better documented.

The bottom line

The 97% statistic is not a claim that AI access control is the only thing that matters. It is evidence that, for the organizations that reported an AI-related breach, the basic identity layer was missing. Approval is day zero. Access review is the recurring work that keeps the approved state and the actual state from drifting apart.

Tags: AI Security Identity and Access IBM Anthropic Claude Shadow AI Access Control

Sources & References

SE SecurityXP Editorial Desk
SecurityXP Editorial Desk Vulnerability Research & News Board

Automated and analyst-reviewed threat intelligence briefings tracking active exploitation campaigns, CVE disclosures, and extortion group activity.

Security Digest

The most important cybersecurity stories, CVE alerts, and threat intelligence — one concise email. Unsubscribe anytime.

Related Articles

Process diagram of the data breach path: Employee access through Legal review, with the fault at 30-day retention.
AI/ML Security

Microsoft restricts employee Claude Fable 5 access over Anthropic data retention

Microsoft restricts employee access to Claude Fable 5 while legal reviews Anthropic's 30-day retention policy, which can retain flagged content for two years.

Process diagram of the software supply chain: Model Access through Workflow Shift, with the fault at Commerce Order.
AI/ML Security

Washington Pulled the Plug on Anthropic ‘s Fable 5 and Mythos 5 models. The Rest of the World Is Watching.

The organizations that had integrated these models into security operations, threat hunting pipelines, and vulnerability research workflows are now running...

Process diagram of the data breach path: Leaked prompts through Data exfiltration, with the fault at Leaked prompts.
AI/ML Security

AI-Enabled Hacker Caught: Leaked Prompts Expose Resume and IP Address

A fully AI-enabled hacker was caught, revealing his full system prompts, which included his resume and his IP address.

Process diagram of the exploitation chain: AI Systems through Annual Audit, with the fault at Security Controls.
AI/ML Security

New EU AI Security Regulations for Organizations

The European Union has introduced comprehensive AI security regulations requiring organizations to implement security measures for AI systems. We break down the requirements, timelines, and compliance steps.