There is a specific kind of organizational risk that is harder to manage than a known vulnerability. It is the risk created by the gap between what security teams believe they can do and what they can actually execute under pressure. The Semperis global study of 1,100 organizations, examining how AI is reshaping identity attack surfaces across eight countries, documents that gap with uncomfortable precision.
Ninety-three percent of organizations already use or plan to use AI agents for sensitive security tasks including password resets and VPN access. Only 32 percent are very confident they could regain control of their identity infrastructure if an AI agent exposed administrative credentials. In France, that confidence figure drops to 12 percent.
That arithmetic describes an industry collectively accelerating into a risk posture it has not yet built the recovery capability to manage. Organizations are not simply underinvesting in AI identity governance. Many are doing so while holding an inaccurate picture of their own resilience posture, which is a considerably more dangerous condition than acknowledged underinvestment.
As enterprises uncover governance blind spots around AI identities, many face a parallel challenge: critical vendor obligations, compliance terms, and operational risk buried inside static contracts. Agiloft CLM + AI transforms disconnected agreements into actionable intelligence for stronger enterprise governance and faster risk-aware decision-making.
What the Study Actually Measured and Why the Scope Matters
The Semperis State of Identity Security in the AI Era study draws on responses from 1,100 organizations across the United States, United Kingdom, France, Germany, Spain, Italy, Singapore, and Australia, covering identity systems including Active Directory, Entra ID, and Okta. The geographic breadth matters analytically because it surfaces meaningful variance in how organizations across different regulatory and cultural environments are approaching AI identity governance.
The country-level divergence in recovery confidence is one of the most telling findings in the dataset. US organizations sit at 53 percent confidence in their ability to regain control following an AI-related identity breach. French organizations sit at 12 percent. That is not a minor variance attributable to survey methodology. It suggests materially different levels of identity resilience investment, recovery planning maturity, and organizational awareness of what an identity breach at the AI agent layer would actually require to remediate.
For global enterprises managing identity infrastructure across multiple geographies, those country-level differences carry direct implications for how recovery plans should be stress-tested and where additional investment is most urgently needed.
AI Is Already Inside the Identity Perimeter. The Governance Has Not Caught Up.
The study’s most operationally significant finding is not about future risk. It is about the current state of AI deployment within identity-sensitive environments.
Ninety-two percent of respondents report that some portion of their workforce has AI installed on local machines where it can access SSH keys and encryption keys. Twenty-nine percent already use AI agents to manage security-related help desk functions including password resets and VPN access. Another 65 percent intend to do so within the next year.
These are not edge cases or experimental deployments. They represent mainstream enterprise adoption of AI agents in roles that carry direct access to credential infrastructure, authentication systems, and the cryptographic material that underpins secure communications across the organization.
The governance picture sitting alongside that deployment reality is considerably less mature. Globally, only 65 percent of organizations report that AI identities are fully registered, authenticated, and authorized within a formal identity management system. Six percent admit they do not track AI identities at all. In organizations that do track them, 57 percent manage AI identities within the same system as human identities, while 43 percent use a separate authentication and authorization framework.
The Unregistered Agent Problem
The 35 percent of organizations where AI identities are not fully registered, authenticated, and authorized in a formal system represents a substantial unmanaged attack surface. An AI agent operating with credentials that are not formally tracked in the identity fabric cannot be governed through standard access review processes, cannot be included in least-privilege enforcement programs, and cannot be efficiently revoked if its credentials are compromised.
That last point is particularly consequential given the detection timelines documented in parallel research from Akeyless, which found organizations take an average of 14 hours to detect a compromised AI agent. An unregistered agent operating with untracked credentials during a 14-hour detection window, with access to SSH keys and encryption material, represents a breach scenario that few existing recovery playbooks are designed to address.
Identity Infrastructure as the AI Breach Blast Radius
Chris Inglis, the first US National Cyber Director and a Semperis strategic advisor, frames the core risk in terms that move beyond technical vulnerability assessment into business consequence territory.
Identity failures, he observes, turn technical incidents into prolonged business crises. The gap between perceived resilience and actual recovery capability is particularly dangerous at the identity layer because identity infrastructure is not simply one system among many. It is the governance fabric that controls access to every other system. When Active Directory or Entra ID is compromised, the blast radius extends to every resource, application, and workflow that depends on those systems for authentication and authorization.
Introducing AI agents into that fabric without commensurate investment in monitoring, recovery planning, and resilience testing does not simply add new attack vectors. It potentially amplifies the blast radius of any identity compromise by introducing actors that operate at machine speed, hold broad credential access, and may not be visible in standard identity audit processes.
The study finding that 74 percent of organizations believe AI will increase attacks on identity infrastructure sits in instructive tension with the recovery confidence figures. Most organizations anticipate the threat. Far fewer have built the recovery capability that their own threat anticipation implies they need.
The Recovery Readiness Problem Is Distinct From the Prevention Problem
Enterprise identity security investment has historically been weighted toward prevention: access controls, multi-factor authentication, privileged access management, and identity governance programs designed to minimize the probability of a breach. Recovery readiness, the ability to restore identity infrastructure to a trustworthy state quickly following a compromise, has received considerably less investment and organizational attention.
That imbalance was a manageable risk in environments where identity systems were relatively stable and breach scenarios were well-understood. It becomes a more serious structural problem when AI agents are being granted access to sensitive identity infrastructure at the pace the Semperis study documents.
The reason is execution speed. An AI agent with access to administrative credentials and encryption keys can move, exfiltrate, and modify at a velocity that makes prevention the only viable defense in traditional security models. But prevention at 100 percent effectiveness is not a realistic security posture. Recovery readiness, the ability to detect, contain, and restore quickly when prevention fails, is not a backup strategy. It is the primary defense mechanism against the worst-case consequences of an AI identity breach.
Grace Cassy of Ten Eleven Ventures identifies observability and recovery readiness as the missing dimensions in most organizations’ AI identity governance programs. That framing reflects a maturing understanding of what AI-era identity security actually requires: not simply controls that prevent bad outcomes, but visibility that enables rapid detection and recovery infrastructure that makes restoration achievable within a timeframe that limits business impact.
Governance Practices That Reflect Current Deployment Reality
The study’s governance recommendations reflect a recognition that the AI identity problem requires adapting existing identity security frameworks rather than waiting for entirely new ones.
Treating AI agents explicitly as non-human identities within the identity fabric is foundational. Organizations that manage AI agent credentials informally or within human identity systems without differentiation lose the visibility needed to apply appropriate governance controls. AI agents behave differently from human users, act continuously rather than in defined sessions, and carry risk profiles that standard user behavior analytics were not designed to detect.
Enforcing least-privilege, just-enough, and just-in-time access for AI agents with the same rigor applied to human privileged access is the control discipline that most directly addresses the credential exposure risk the study documents. An AI agent provisioned with only the access required for its specific task at the moment of execution, with credentials that expire immediately upon task completion, carries a fundamentally different risk profile than one operating with persistent broad credentials.
The recommendation to use user and entity behavior analytics to detect anomalous or dormant agent behavior addresses a specific failure mode that the study implies is already occurring: AI agents that remain active with valid credentials beyond their intended use case, accumulating access that is never reviewed or revoked. Detecting these effectively requires analytics systems calibrated for non-human identity behavior patterns, which most existing UEBA deployments have not been configured to address.
The 83 Percent Priority Signal and What It Means for Vendor Positioning
The finding that 83 percent of respondents identify AI identity governance as a priority for the coming months is the most commercially significant data point in the study for vendors operating in the identity security category.
Priority at that scale, documented across 1,100 organizations in eight countries, describes active buying intent rather than aspirational planning. Organizations that have identified AI identity governance as a priority and simultaneously acknowledge they are not confident in their current recovery capability are describing a security gap with a budget allocation attached to it.
The identity security vendors best positioned to capture that intent are those that can credibly address both dimensions of the problem: governance controls that bring AI agent identities into managed, auditable infrastructure, and recovery capabilities that address the resilience gap the study documents at the organizational level. Point solutions that address one dimension without the other will face increasing pressure in evaluation processes where the Semperis data is being used as a benchmark for assessing organizational readiness.
For security leaders building the internal case for AI identity governance investment, the study provides exactly the kind of peer benchmark data that moves budget conversations from security team advocacy to executive-level risk acknowledgment. The combination of deployment scale, governance maturity gaps, and recovery confidence figures documented across 1,100 organizations gives CISOs a credible external reference point for investment justification that does not rely solely on internal threat assessments.
The organizations that act on that 83 percent priority signal with concrete investment in the next two quarters will be meaningfully better positioned than those that treat it as a planning aspiration for a future budget cycle. The AI agents are already inside the identity perimeter. The governance clock is running.
Research and Intelligence Sources: Semperis
To participate in our interviews, please write to our CyberTech Media Room at info@intentamplify.com
