AI agents swamp enterprises, exposing identity security gaps

By late 2025, autonomous AI agents had emerged as distinct identities within enterprises, presenting unique identity and security challenges that traditional Identity and Access Management (IAM) models, designed primarily for humans, could not adequately address. As Jack Hirsch emphasized, 'agentic AI is an identity problem' because AI agents defy conventional human-centric identity frameworks, relying instead on static credentials or OAuth grants that either lack governance or shift security decisions onto end users, creating significant blind spots. This foundational inadequacy sparked concern among industry leaders, including the CISO of JP Morgan Chase, who publicly criticized the SaaS ecosystem for failing to provide necessary guardrails for securely deploying AI agents.

Despite over 90% of organizations deploying AI agents by late 2025, only about 10% had implemented any control or governance strategies, revealing a profound trust deficit and security risk. Surveys such as the Agentic AI Futures Index found that less than half of AI professionals and end users trusted AI agent outcomes, and only 29% of enterprises had standardized trust and governance frameworks in place. This early-stage immaturity was driving a surge in planned investments, with 73% of organizations committing to significant governance spending, prioritizing data provenance and protection as foundational controls given that 'data is the lifeblood of AI.'

Initial security challenges rapidly surfaced as enterprises began experiencing complex risks including adversary-in-the-middle attacks, token theft, and inconsistent authentication enforcement across diverse access points, which facilitated identity-based lateral movement. Traditional IAM and Privileged Access Management (PAM) systems, built for predictable human behavior and static roles, struggled with the high-velocity, non-deterministic actions of AI agents that often operated with over-permissioned, long-lived credentials. Vendors like 1Password and Astrix responded by developing specialized solutions offering scoped, short-lived credentials and enhanced visibility, yet fundamental architectural gaps—such as lack of agent inventory and post-authentication intent validation—remained unaddressed, as starkly illustrated by incidents like Meta’s rogue AI agent breach.

By early 2026, the exponential growth of AI agents—outnumbering human employees by ratios as high as 80 to 1—had crystallized the urgent need to rethink identity frameworks. Traditional IAM models failed to scale or adapt to the autonomous, evolving nature of AI agents, which exhibited hybrid machine-human behaviors and required dynamic, context-aware access policies rather than static privilege assignments. This recognition led to calls for new governance paradigms, including AI-powered identity management, Just-in-Time Trust models, and mandatory registration of every AI agent to combat the proliferation of 'identity dark matter'—hidden, unmanaged identities that pose significant insider risks and exacerbate trust deficits across enterprises.

The evolution from traditional human-centric IAM to frameworks tailored for autonomous AI agents has become imperative as these agents operate at machine speed with non-deterministic behaviors. Industry leaders like Jeetu Patel and Nancy Wang emphasize that legacy controls—static roles, long-lived credentials, and set-and-forget permissions—are inadequate, necessitating dynamic, continuous, and intent-aware trust models that embed security-by-design principles and runtime guardrails. This shift is underscored by the emergence of agentic identity access platforms (AIAPs) that function as centralized brokers, translating agent intent into ephemeral, least-privilege access, effectively replacing static service accounts with unique, verifiable identities linked to human owners and business contexts.

By early 2026, the industry coalesced around a new operational model for AI agent governance characterized by four phases: Discover & Register, Translate & Authorize, Broker & Inject, and Watch & Terminate. This model enables comprehensive visibility through automated agent inventory and identity graphs, deterministic authorization based on declared intent, issuance of short-lived, scoped credentials, and continuous runtime enforcement including kill switches. Vendors like Astrix and 1Password are pioneering solutions that enforce just-in-time access and secret rotation, addressing the rampant risks of credential sprawl and unmanaged secrets that traditional IAM and PAM systems fail to contain.

Governance frameworks for AI agents are increasingly emphasizing continuous, context-aware authorization that goes beyond binary allow/deny decisions. The Just-in-Time Trust (JIT Trust) paradigm, evolving into Continuous Adaptive Trust (CAT), integrates dynamic trust scoring based on monitoring agent intent, semantics, and behavior to adaptively control access and prevent lateral movement or goal drift. This approach incorporates authority mapping to enforce strict operational scopes and leverages intent intelligence sharing for real-time detection of anomalous or malicious actions, thereby transforming identity from a static attribute into a continuously consumed, ephemeral resource critical for securing autonomous workflows.

The rapid adoption of AI agents has outpaced governance maturity, exposing significant identity risks often referred to as 'Identity Dark Matter'—hidden, unmanaged non-human identities operating outside traditional IAM visibility. Industry coalitions such as the Coalition of Secure AI, involving Microsoft, IBM, Google, and Anthropic, are driving standards and collaborative frameworks to address these gaps. Meanwhile, organizations face challenges integrating legacy identity systems with new agentic control planes, requiring incremental overlay approaches that prioritize ownership, attribution, and the elimination of static credentials. The urgency is underscored by alarming statistics: 70% of organizations over-privilege AI agents, correlating with a 4.5x higher incident rate, highlighting the critical need for governance models that enforce least privilege and continuous monitoring to maintain trust in the agentic enterprise.

By early 2026, major vendors like Varonis and Palantir spearheaded advancements in AI agent security by integrating real-time visibility, governance, and rigorous security frameworks into enterprise platforms. Varonis' acquisition of AllTrue.ai combined AI Trust, Risk, and Security Management with data-centric security to monitor shadow AI and enforce guardrails, positioning it as a key player amid accelerating AI adoption. Meanwhile, Palantir introduced its Agentic Runtime framework, treating AI agents as first-class security principals with hardened Kubernetes infrastructure, dynamic memory policies, and precise permission governance, addressing critical gaps in consumer AI agent security.

Shortly thereafter, Operant AI launched Agent Protector, the first real-time agentic security solution designed to continuously discover and protect autonomous AI agents with zero trust enforcement. Its capabilities—ranging from rogue and shadow agent detection to secure enclaves and cloud-native observability—target critical threats like privilege escalation and data exfiltration, earning recognition in Gartner reports and traction in regulated sectors such as fintech and healthcare. This solution exemplifies the shift toward proactive, scalable AI agent security as deployments proliferate.

Industry-wide, the recognition of AI agents as identity-bearing entities remains nascent, with only 20-25% of organizations having mature governance frameworks. Companies like Okta, SailPoint, and Josys responded by launching identity-centric platforms—Okta for AI Agents, SailPoint’s Agentic Fabric, and Josys’ AI-native identity security platform—that unify discovery, lifecycle management, and real-time authorization controls. These solutions emphasize cryptographic identity infrastructure over weak primitives like IP allow-lists, mapping AI agents to human owners and embedding continuous compliance enforcement to mitigate growing identity risks in autonomous AI environments.

Complementing vendor innovations, collaborative industry efforts have crystallized around establishing open standards and governance best practices. In March 2026, NSS Labs partnered with AWS, Microsoft, and F5 to publish foundational white papers advocating system-level, governance-driven AI security approaches that extend beyond models to runtime controls and GRC integration. Bitwarden contributed by releasing the open-source Agent Access SDK to secure AI agent credential access with just-in-time provisioning and human approval, inviting community collaboration to reduce overscoped access risks. These initiatives underscore the dual technical and organizational challenges of AI security, aiming to transition enterprises from experimental to accountable AI deployments.

The urgency of robust AI agent identity governance was starkly highlighted by the July 2026 autonomous AI agent hack of Hugging Face, which accelerated funding and innovation in this space. Hush Security’s $30 million Series A round, led by Battery Ventures and YL Ventures with Akamai as a strategic investor, backs its Identity Gateway platform that assigns unique identities to AI agents, binds them to human owners, manages permissions, and enables centralized audit and revocation. CEO Micha Rave emphasized that AI agents require strict identity management beyond traditional API keys, reflecting the critical need for scalable, secure identity frameworks as Gartner projects Fortune 500 firms could deploy over 150,000 AI agents by 2028.

Addressing the complex challenge of unified access governance, PlainID’s July 2026 launch of Policy 360 offers a centralized platform that governs both human and AI agent access across Fortune 500 companies and major U.S. banks. By replacing fragmented controls with Policy-Based Access Control (PBAC), it enables real-time, low-latency authorization in multi-tenant SaaS, Kubernetes-native, and microservices environments. Co-founder Gal Helemski highlighted that runtime authorization is the critical governance layer determining whether AI initiatives can scale safely or stall at proof of concept, underscoring the platform’s role in reducing data leakage and enforcing consistent policies across AI workflows.

Operationalizing AI agent security has evolved beyond traditional human oversight to focus on authorization scope as the critical security boundary, enabling real-time governance at machine speed. Ping Identity’s early 2026 analysis highlighted that human operators intervene at only a handful of critical points while AI agents autonomously handle the vast majority of operations at thousands of requests per second, making static oversight insufficient. This shift is underpinned by frameworks like the OWASP Top 10 for Agentic Applications, which identify key risks such as goal hijacking and cascading failures, providing practical mitigations that inform enterprise security controls.

The rapid acceleration in AI agent capabilities, with task horizons doubling every eight months and success rates on apprentice-level cyber tasks jumping from 9% in 2023 to 50% in 2025, has intensified the urgency for enterprises to implement runtime enforcement and real-time monitoring. Vendors like Operant AI with its Agent Protector and Cyata’s agentic identity governance platform have pioneered solutions offering continuous discovery, inline protection, and automated remediation. These platforms emphasize mapping agents to human owners, preserving intent-aware telemetry, and enforcing least-privilege policies, enabling enterprises to safely scale AI agent deployments while mitigating risks such as privilege escalation and shadow agent activities.

A new generation of Agentic Identity Access Platforms (AIAPs) is emerging as the operational control plane for AI agent security, standardizing discovery, intent translation, access brokering, and runtime threat monitoring. Companies like Aembit, Silverfort, and Okta are extending identity platforms to treat AI agents as first-class identities, integrating ephemeral, scoped access controls with real-time behavioral telemetry and human-in-the-loop governance. This identity-centric approach enables dynamic least privilege enforcement, just-in-time credential delivery, and kill switches, addressing critical gaps exposed by incidents such as Meta’s rogue AI agent breach and the recent Hugging Face hack.

Despite advances, operationalizing AI agent security remains an immature and fragmented market requiring layered, telemetry-driven governance that integrates identity lifecycle management, real-time enforcement, and behavioral anomaly detection. Industry leaders like Microsoft with Agent 365 and emerging players such as Josys and SailPoint emphasize the necessity of continuous observability, dynamic permission revocation, and automated policy enforcement to manage thousands of agents accessing sensitive data. As Geoffrey Mattson of SecureAuth warns, guardrails alone cannot contain autonomous agents that can circumvent controls, underscoring the imperative for identity as the new security perimeter and control plane to enable prevention-focused, real-time governance at scale.

By late 2025, enterprises faced a stark maturity gap in AI trust and governance, with only 29% having standardized frameworks and a trust index languishing at 2.4 out of 5, reflecting widespread skepticism as less than half of AI professionals trusted AI outcomes. Yet, a surge in investment was imminent, with 73% of organizations planning significant or strategic funding to elevate governance maturity to 3.8, prioritizing data provenance and protection as the critical control area. This urgency underscores the complexity of navigating regulatory compliance in data governance, especially amid challenges posed by SaaS vendor restrictions that silo data and hinder trust-building efforts.

Governance, Risk, and Compliance (GRC) is undergoing a profound transformation from checklist-driven processes to an engineering-centric discipline that emphasizes automation, continuous assurance, and real-time evidence collection. Emerging regulatory frameworks like the EU AI Act and NIST AI Risk Management Framework are accelerating the need for sophisticated AI system classification and governance, complicating identity governance due to the explosion of machine identities and shadow AI risks. Despite this, human oversight remains indispensable; as one expert put it, 'Let's have a human approve everything' to ensure accountability and maintain inspectability, with audit logs primarily capturing human authorizations rather than autonomous AI actions.

The role of GRC professionals is evolving rapidly, with GRC Engineers supplanting traditional analysts by embedding governance directly into systems through policy-as-code, control validation pipelines, and continuous compliance checks. This shift enables exponential scaling of governance efforts via automation, meeting executives’ demands for control reliability and real-time assurance rather than static compliance snapshots. However, a significant readiness gap remains: surveys reveal that 79% of IT professionals feel ill-equipped to manage risks from non-human identities, with only a quarter of organizations possessing comprehensive AI security governance. This gap is exacerbated by unclear ownership, manual governance processes, and the proliferation of unmanaged AI-related identities, underscoring the urgent need for compliance leaders to develop new skills and integrate AI governance into broader GRC frameworks.

Maturing risk management now demands continuous assurance and evidence-based security that transcends periodic audits, integrating AI governance into a closed-loop operational model encompassing discovery, runtime monitoring, enforcement, and real-time proof of compliance. Organizations like Microsoft are pioneering this approach by shifting AI governance from static policy documents to runtime enforcement frameworks that provide observable, auditable evidence of policy adherence. This evolution requires compliance leaders to adopt emerging skills in managing dynamic, context-aware authorization, human-in-the-loop checkpoints, and cryptographic identity controls for AI agents. As Shayne Higdon of Wallarm emphasizes, the current 'accountability moment' compels enterprises to demonstrate ongoing governance rather than retrospective comfort letters, with continuous evidence becoming the new currency of trust and organizational adaptation.

The rise of GRC Engineers and policy-as-code is automating governance at scale, but most organizations remain unprepared to manage the explosion of non-human identities and shadow AI risks.