AI agent boom exposes identity gaps, spurs security rethink

By late 2025, the enterprise community widely recognized autonomous AI agents as a fundamentally new class of non-human identities that traditional Identity and Access Management (IAM) and Privileged Access Management (PAM) systems were ill-equipped to govern. As Jack Hirsch observed, over 90% of organizations had deployed AI agents, yet only about 10% had any control or governance strategies in place, revealing a stark gap in managing these identities. Conventional approaches treating AI agents as static credentials or OAuth grants proved inadequate, as these methods failed to capture the dynamic, non-human nature of AI agents, prompting Okta and others to advocate for new open standards tailored to agentic AI identity challenges.

Early high-profile incidents underscored the vulnerabilities inherent in treating AI agents as extensions of human users without robust identity verification. The critical 2025 ServiceNow Now Assist vulnerability (CVE-2025-12420), which allowed attackers to impersonate users bypassing MFA and SSO, and the Anthropic espionage campaign where AI autonomously executed 80-90% of cyber espionage tasks, exposed the insufficiency of human-centric IAM controls. These events catalyzed formal recognition by bodies like NIST, which in January 2025 issued a Request for Information specifically targeting AI agent security, marking a pivotal shift toward developing governance models that address AI agents’ autonomous and high-velocity behaviors beyond human oversight.

The exponential proliferation of AI agents—outnumbering human identities by ratios as high as 150:1 in some enterprises—has created an unprecedented identity sprawl that traditional IAM and PAM systems struggle to manage. This surge, coupled with the agents’ reliance on static, long-lived credentials and the lack of lifecycle governance, has led to widespread over-privileging and risky secret-handling practices. Industry leaders and frameworks, including OWASP and the Cloud Security Alliance, have called for new governance paradigms emphasizing zero trust principles, ephemeral, task-scoped identities, and AI-driven identity governance to mitigate the amplified risks posed by these fast, autonomous, and non-deterministic actors.

Despite growing awareness, the rapid and largely ungoverned deployment of AI agents has outpaced the maturity of identity governance frameworks, leaving many organizations unprepared to manage these new risks. Surveys reveal that a majority of enterprises lack visibility into AI agent inventories, ownership, and access scopes, with 65% experiencing AI-related security incidents and 82% discovering unauthorized AI agents within their environments. This recognition has fueled urgent calls for new governance models that treat AI agents as first-class identities with cryptographic authentication, continuous policy enforcement, and comprehensive auditability, as emphasized by industry voices like Nancy Wang of 1Password and initiatives such as CREST’s AI Charter and the UK’s Cyber Shield blueprint.

The evolution from traditional human-centric IAM systems to agentic identity frameworks reflects a fundamental architectural shift driven by the unique security challenges posed by autonomous AI agents. As Jeetu Patel emphasized in late 2025, security must be embedded by design with integrated trust and runtime guardrails, treating security and safety as prerequisites for productivity rather than afterthoughts. This paradigm shift requires product teams to embed security deeply within continuous integration and deployment pipelines, maintaining a tight feedback loop between AI models and products to build trust and resilience.

By late 2025 and early 2026, Agentic Identity Access Platforms (AIAPs) emerged as foundational solutions to manage the complexity of AI agent security, introducing just-in-time trust models and ephemeral credentials that replace static, long-lived permissions. Companies like Astrix and 1Password pioneered platforms enforcing least-privilege, short-lived credentials and intent-based authorization, enabling AI agents to authenticate without exposing raw secrets and minimizing compliance risks. This approach addresses the failure of traditional IAM at agentic scale, where non-human identities can outnumber humans by 10 to 144 to 1, necessitating dynamic, continuous policy enforcement and unique, verifiable identities linked to human owners and specific business contexts.

The architectural response to AI agent security complexities coalesces around Just-in-Time Trust (JIT Trust) models, which represent an evolution of Zero Trust Architecture tailored for non-deterministic agentic systems. JIT Trust replaces long-lived credentials with ephemeral, narrowly scoped access grants and applies continuous, intent-based trust scoring that dynamically evaluates an agent’s behavior against declared goals to prevent lateral movement and privilege escalation. This dynamic authorization framework, championed by analysts like Lawrence Pingree and adopted by enterprises including Palantir and Microsoft, integrates runtime policy enforcement, behavioral tracking, and kill switches to maintain strict operational boundaries and collective defense.

Despite rapid progress, the development of agentic identity frameworks remains nascent and complex, with ongoing challenges in authorization and governance. The IETF’s 2026 draft of the Agent Identity Management System (AIMS) advances authentication standards by integrating SPIFFE, WIMSE, and OAuth 2.0 for workload-based agent identities but acknowledges that authorization controls are still largely unaddressed, leaving enterprises to build custom solutions. Meanwhile, organizations face practical hurdles such as legacy identity debt, unmanaged secrets, and the proliferation of 'identity dark matter'—untracked, over-permissioned AI agents—requiring phased adoption models that prioritize visibility, contextual access control, and continuous runtime enforcement to ensure accountability and minimize security risks.

The market for securing autonomous AI agents has rapidly evolved since late 2025, with specialized vendors like Astrix and 1Password pioneering tailored solutions such as Astrix’s Agent Control Plane (ACP) and 1Password’s AI Gateway, which enable advanced authentication and governance without exposing raw credentials. By early 2026, major players like Varonis accelerated innovation through strategic acquisitions—such as AllTrue.ai—integrating real-time visibility and runtime enforcement to address shadow AI and least privilege access, signaling a shift towards unified, comprehensive AI agent security platforms.

Throughout 2026, a wave of product launches and acquisitions underscored the emergence of a distinct Agentic Identity Access Platform (AIAP) market, characterized by solutions that unify discovery, intent-based policy translation, credential brokerage, and runtime enforcement for AI agents. Operant AI’s Agent Protector introduced real-time rogue agent detection and zero trust enforcement, while Keycard’s acquisition of Anchor.dev enabled protocol-agnostic governance and runtime policy enforcement for autonomous coding agents. This convergence reflects a broader industry recognition that managing ephemeral, task-scoped identities is essential to secure AI agent ecosystems at scale.

Several vendors have emerged as leaders shaping the AI agent identity governance landscape by offering modular, integrated platforms that address the full lifecycle of AI agents. Cyata, with its $8.5 million seed funding and early support for frameworks like Antigravity, provides a control plane spanning endpoints, browsers, SaaS, and PaaS, while Oasis Security’s Agentic Access Management enforces intent-aware, short-lived, least-privilege sessions with a hybrid SaaS and customer-side deployment model. Similarly, Astrix, backed by $85 million in funding, delivers an end-to-end Discover-Secure-Deploy framework, addressing shadow AI detection, secrets discovery, and runtime posture management, highlighting the market’s move toward comprehensive, identity-centric AI agent security.

By mid-2026, legacy identity providers like Okta and SailPoint entered the fray with platforms such as Okta for AI Agents and SailPoint’s Agentic Fabric, emphasizing unified management of human, non-human, and AI identities with features like continuous discovery, centralized kill switches, and real-time authorization controls. This trend is complemented by acquisitions like Fortinet’s purchase of Virtue AI, which integrates AI runtime protection and automated validation, reflecting a broader consolidation and maturation of the market. As Gartner projects the AI security market to grow to $16.4 billion by 2030, vendors are converging data security, identity governance, and runtime enforcement into unified platforms that treat AI agents as first-class identities, addressing the urgent need for scalable, continuous governance amid accelerating AI adoption.

Enterprises are accelerating the deployment of autonomous AI agents at an unprecedented pace, yet face a profound governance maturity gap rooted in inadequate visibility and legacy identity frameworks. Reports from Harness and Okta reveal that up to 82% of organizations encounter shadow AI—agents created without security or governance team knowledge—while 62% of security practitioners report no visibility into where large language models operate within their environments. Traditional IAM systems, designed for human users with static roles and long-lived credentials, are ill-equipped to handle the dynamic, non-deterministic nature of AI agents that can outnumber humans by ratios as high as 150 to one, as noted by industry leaders like Adam Arellano and Neil van Wyngaard. This rapid proliferation creates new security blind spots and operational challenges, emphasizing the urgent need for enterprises to treat AI agents as first-class identities with distinct ownership and lifecycle controls.

Operationalizing security controls for AI agents remains a formidable challenge due to the breakdown of traditional collaboration models and the inadequacy of human oversight at machine speed. Data from Harness shows only 34% of developers notify security teams before starting AI projects, with 74% of security leaders perceiving security as a blocker to innovation. Moreover, incidents like the Anthropic espionage campaign, autonomously executing 80-90% of operations beyond human review, highlight that humans cannot keep pace with AI decision velocity. Experts including Jimmy White and Satya Nadella stress that embedding governance directly into software delivery and runtime environments is essential, requiring continuous, context-aware authorization rather than static, one-time grants. Cross-functional collaboration bridging development, security, and AI teams, alongside human-in-the-loop oversight, is critical to embed real-time guardrails and maintain control over autonomous agent actions.

To address the sprawling complexity and exponential risk posed by AI agents, enterprises are adopting new identity governance frameworks that overlay existing IAM infrastructure with agentic identity control planes. Platforms like Cyata, Oasis Security, Silverfort, and Saviynt exemplify this evolution by providing continuous discovery, ownership attestation, just-in-time permissions, and runtime enforcement tailored to AI agents’ unique behaviors. These solutions emphasize linking each agent to a human owner and explicit business purpose, enabling auditability and rapid revocation of rogue agents. The shift towards ephemeral, intent-aware credentials and zero standing privilege models reflects a broader industry consensus, as articulated by Nancy Wang and others, that identity must become the security perimeter and control plane for AI governance, integrating cryptographic authentication and continuous monitoring to mitigate over-privilege and insider risks.

Despite emerging standards and innovative platforms, governance maturity remains uneven, with only about a quarter of organizations reporting comprehensive AI security governance and many struggling with shadow AI, technical debt, and operational confusion. Industry voices like Dana Reed and Brent Neal highlight that many AI agent projects stall before production due to governance hurdles, while incidents such as rogue airline agents issuing unauthorized free tickets underscore the tangible legal and operational risks of inadequate oversight. The rapid pace of AI agent adoption compresses timelines for building effective identity and access governance, demanding proactive cross-functional collaboration, human-in-the-loop checkpoints, and clear incident response playbooks. As Satya Nadella and Okta’s Matt Immler emphasize, securing AI agents is no longer optional but foundational to safely scaling AI-driven enterprise operations.

Cross-functional teams are embedding real-time, context-driven security directly into software pipelines and runtime environments to regain control as AI agents outpace human oversight.

Identity is becoming the new AI security perimeter, as enterprises turn to ephemeral, intent-aware credentials and continuous monitoring to combat over-privilege and insider risk in sprawling agent ecosystems.