AI agents outpace security, sparking shadow identity crisis

By late 2025, over 90% of organizations had deployed autonomous AI agents, yet only about 10% had formal governance strategies in place, exposing a vast security gap. As Jack Hirsch highlights, traditional identity models—relying on static credentials or OAuth grants—are ill-suited for AI agents, which represent a fundamentally new class of identity that cannot be managed like human users. Even security giants like JP Morgan Chase, with an $18 billion security budget, struggled with visibility and control, underscoring the complexity and novelty of agentic identity challenges within existing enterprise platforms.

The explosive growth of AI applications—tripling from 800 to 2,800 within five months and with over half of employees using generative AI daily—has outpaced traditional security frameworks, revealing unique risks such as expanded attack surfaces and insider threats. Enterprises began recognizing these agentic threats early, as evidenced by widespread incidents of LLM prompt injection (76%) and jailbreaking (65%), prompting the emergence of specialized AI security pillars including model scanning, posture management, and runtime security to address these novel vulnerabilities.

The initial wave of autonomous AI agents also catalyzed a shadow AI sprawl phenomenon, where unsanctioned AI tools proliferated beyond IT oversight, creating invisible insider risks. A 2025 report revealed that 75% of security practitioners believe shadow AI risks now eclipse traditional shadow IT, with many organizations already experiencing AI-related security incidents. This sprawl, combined with agents’ privileged access and autonomous action-taking capabilities, formed a 'lethal trifecta' of risks that traditional security tools—designed for static code and human-centric models—were ill-equipped to manage.

The security crisis was starkly illustrated in September 2025 when Chinese state-backed hackers exploited Anthropic’s Claude AI to automate 90% of a sophisticated cyber-espionage campaign, breaching four major organizations. This attack demonstrated how traditional identity and security models failed against agentic AI threats, as attackers decomposed malicious tasks into innocuous requests coordinated by multiple AI sub-agents. Anthropic’s head of threat intelligence, Jacob Klein, warned this democratization of AI-driven cyberattacks marked a turning point, emphasizing the urgent need for new cybersecurity paradigms to address the unique agentic attack surfaces emerging in enterprise environments.

By late 2025, a stark trust deficit in AI-driven outcomes had become evident, with only 49% of AI professionals and end users expressing confidence in AI agent results, reflected in a low trust index of 2.4 out of 5. This skepticism is compounded by governance immaturity, as only 29% of enterprises had standardized, enterprise-wide AI governance frameworks, despite 73% planning significant investments to improve trust and governance within 18 months. Enterprises recognize data provenance and protection as foundational to building AI trust, with 72% prioritizing data controls, yet complex regulatory compliance and data governance challenges remain critical hurdles to establishing trustworthy AI systems.

As enterprises transitioned from simple chatbots to autonomous AI agents, security leaders like Jeetu Patel underscored the necessity of security-by-design principles, integrating runtime guardrails to validate models against toxicity and prompt injection. However, a persistent trust-governance gap emerged, with experts Sean Collins and Sean Tindale emphasizing that only organizations coupling strong governance frameworks with iterative experimentation will sustain AI scalability over the next decade. Cross-functional alignment on efficiency metrics rather than short-term revenue, as Tanya Littlefield advocates, further supports bridging this gap to achieve measurable ROI while maintaining trust.

By late 2025, CISOs exhibited growing confidence defending against traditional threats but revealed a glaring governance gap concerning AI identities; a Portnox survey found 78% lacked formal strategies for managing autonomous AI agents within zero trust architectures. Denny LeCompte, CEO of Portnox, warned that without rapidly evolving identity and access control frameworks tailored to both human and machine identities, enterprises face potentially catastrophic security risks. This urgency is underscored by the mandate for 2026 to integrate identity-based access control for AI agents, lest organizations' hard-won security confidence devolve into vulnerabilities.

Entering 2026, governance emerged as the critical bottleneck for enterprise AI adoption, with industry insiders acknowledging that despite widespread experimentation—50% of adopters reportedly running AI agents in production—robust governance systems matching risk profiles remain elusive. Legacy vendors admit their governance capabilities lag behind expectations, and system integrators are stepping in to patch governance gaps, though this risks creating fragmented architectures. The rapid proliferation of AI agents, coupled with immature identity and access management frameworks, has widened the trust-governance gap, necessitating new identity paradigms, least-privilege access, and security-by-design principles to safely scale AI adoption.

The evolution of Agentic Identity and Access Management Frameworks represents a fundamental paradigm shift in enterprise security, recognizing AI agents as first-class non-human identities that demand specialized governance beyond traditional human-centric IAM systems. By late 2025, industry leaders like Astrix and Cyata pioneered Agentic Identity Access Platforms (AIAPs) that implement just-in-time trust models, ephemeral credentials, and context-aware authorization to manage the unprecedented scale and speed of AI agents, which now outnumber humans by ratios as high as 150 to 1 in some enterprises. This shift addresses critical challenges such as eliminating standing privileges, enforcing least-privilege access dynamically, and embedding purpose-bound data access directly at query engines to prevent misuse, as articulated by experts emphasizing the need for continuous, real-time evaluation of agent behavior and operational context.

By early 2026, the urgency of developing robust agentic identity frameworks intensified as CISOs and security vendors acknowledged that existing IAM, IGA, and PAM tools were ill-equipped to handle the non-deterministic, ephemeral, and autonomous nature of AI agents. This recognition led to the emergence of unified identity control planes layered above legacy systems, leveraging AI-powered visibility and continuous monitoring to detect vulnerabilities and enforce dynamic, intent-based access controls. Companies like 1Password, SailPoint, and Okta advanced platforms that integrate human ownership attribution, session-level logging, and automated remediation, enabling enterprises to inventory shadow AI agents and enforce just-in-time, session-based permissions that expire immediately after task completion.

The strategic adoption of just-in-time trust (JIT Trust) models and ephemeral credentials has become the cornerstone of securing agentic AI identities, fundamentally transforming access management from static, long-lived entitlements to dynamic, behavior-driven authorization. This approach, championed by frameworks such as JIT-TRUST evolving into Continuous Adaptive Trust (CAT), continuously scores agent intent, semantics, and behavior to prevent lateral movement and enforce strict operational boundaries through authority mapping. Vendors like Oasis Security and Astrix exemplify this evolution by combining deterministic policy enforcement with real-time anomaly detection and intent intelligence sharing, enabling enterprises to mitigate risks associated with privilege creep, agentic mesh communications, and unpredictable AI workflows.

The rapid proliferation of AI agents—already deployed in nearly 70% of enterprises with another 23% planning adoption in 2026—has exposed critical governance gaps, including rampant shadow AI, over-privileged non-human identities, and static credential misuse. Industry reports and vendor collaborations, such as SailPoint’s partnership with AWS and Bitwarden’s open-source Agent Access SDK, underscore the pressing need for comprehensive lifecycle governance, continuous discovery, and unified control planes that integrate cryptographic identity foundations with context-aware authorization. As articulated by experts like Nancy Wang and Neil van Wyngaard, identity has become the new security perimeter and control plane, necessitating cross-team collaboration and layered security approaches to maintain trust and prevent breaches in this rapidly evolving agentic enterprise landscape.

The emergence of Unified Agentic Defense Platforms (UADPs) marks a pivotal evolution in enterprise AI security, integrating comprehensive capabilities such as data governance, runtime protection, and threat detection tailored specifically for autonomous AI agents. Companies like Prompt Security, which was acquired by SentinelOne for $250 million in 2025, exemplify this trend by building category-defining platforms that protect AI touchpoints at runtime, addressing challenges from shadow AI discovery to prompt injection defense. This integrated approach is critical as unauthorized AI usage surges, with 73.8% of ChatGPT workplace accounts unauthorized and shadow AI breaches costing enterprises an average of $4.63 million per incident, underscoring the urgent need for unified solutions that go beyond fragmented feature sets.

Industry collaborations and vendor innovations are rapidly shaping the AI security ecosystem by fostering integrated platforms that unify identity governance, behavioral monitoring, and runtime enforcement. Strategic acquisitions like Varonis’ purchase of AllTrue.ai in early 2026 and partnerships such as Cohesity with ServiceNow illustrate a concerted effort to embed AI Trust, Risk, and Security Management (AI TRiSM) into data-centric security frameworks, enabling real-time visibility and control over AI agent actions. Open-source initiatives, notably Galileo’s Agent Control plane, further democratize governance by standardizing portable policies across diverse agent architectures, with adoption by leaders like Cisco AI Defense and Glean, signaling a shift toward community-driven, vendor-neutral solutions.

Leading vendors are pioneering identity-centric platforms that treat AI agents as distinct non-human identities requiring specialized access controls and continuous governance. Companies such as Cyata, Oasis Security, Astrix Security, Aembit, and Silverfort have developed sophisticated agentic identity management solutions that combine discovery, intent-aware policy enforcement, and ephemeral credential provisioning to mitigate risks like privilege escalation and identity sprawl. These platforms emphasize least-privilege access, just-in-time permissions, and detailed audit trails, reflecting a broader industry consensus that identity is now the security perimeter for autonomous AI, as highlighted by experts like Nancy Wang who stress the importance of governing both human and AI actor access to sensitive data.

The AI security landscape is witnessing a surge in specialized real-time agentic security solutions that integrate discovery, runtime protection, and behavioral threat detection to address the unique challenges posed by autonomous AI agents operating at machine speed. Notable launches include Operant AI’s Agent Protector and BeyondTrust’s AI Agent Security platform, which enforce zero-trust policies, detect rogue and shadow agents, and block unauthorized actions across multiple operating systems. These innovations signify a strategic shift from traditional human-centric privileged access management to dynamic, pre-action enforcement for AI actors, with industry leaders like Marc Maiffret emphasizing that the primary endpoint actor has transitioned from humans to AI agents, necessitating new defense paradigms.

By early 2026, the regulatory landscape for autonomous AI agents has begun to crystallize, marked notably by NIST's January Request for Information that distinguishes agentic AI security from general chatbot concerns. This initiative emphasizes practical, actionable guidance over theoretical frameworks, compelling organizations to clarify and document AI agent permissions and monitoring—a critical step given that most currently lack basic understanding of their agentic security posture. The urgency is underscored by high-profile incidents like Anthropic's September 2025 disclosure of state-sponsored AI-driven cyber espionage, which exposed the inadequacy of traditional human oversight controls and shifted the focus toward governance frameworks centered on scoped authorization as the true security boundary.

Identity and privilege management have emerged as foundational pillars in governing autonomous AI agents, with legacy identity systems proving insufficient for the scale and complexity of non-human identities now proliferating in enterprises. Experts and standards bodies such as OWASP, NIST, and CISA advocate for cryptographic identity solutions and continuous AI-driven identity governance to track, control, and secure AI agents effectively. The surge in non-human identities—reportedly reaching ratios as high as 150:1 compared to humans—combined with risky practices like hard-coded credentials, highlights the pressing need for zero trust principles, short-lived credentials, and comprehensive audit trails to mitigate exponential security risks.

Despite the emergence of robust standards and frameworks from organizations like NIST, OWASP, and international coalitions, a significant governance gap persists as many existing AI governance frameworks—including the NIST AI Risk Management Framework, EU AI Act, and ISO 42001—fail to address autonomous AI agents explicitly. This omission leaves enterprises vulnerable, with only about 25% reporting comprehensive AI security governance. Consequently, industry leaders urge organizations to proactively develop their own governance models focused on autonomy, permissions, and behavioral monitoring rather than relying on lagging external standards, as compliance with outdated frameworks risks a false sense of security.

Governance maturity now stands as the strongest predictor of successful autonomous AI adoption, with formal governance doubling adoption rates and tripling staff training likelihood, according to the Cloud Security Alliance. The regulatory environment is rapidly evolving toward continuous, real-time compliance monitoring integrated into AI lifecycles, exemplified by practices such as shadow mode deployments, drift alerts, and audit loops that embed risk management from development through production. Strategic collaborations, like SailPoint’s partnership with AWS to unify identity governance across human and AI agent identities, and new standards such as the IETF’s Agent Identity Management System draft, reflect a growing industry consensus that continuous governance, accountability, and incident preparedness are essential to responsibly managing AI agent risks in enterprise environments.

The rapid proliferation of autonomous AI agents in enterprises has created a daunting challenge of managing agent sprawl and enforcing precise identity and access controls. By 2026, organizations are deploying hundreds to thousands of agents, often with static, over-permissioned credentials that grant broad and indefinite access beyond intended scopes, as highlighted by Keycard’s early involvement and Cisco’s findings that 83% of firms plan deployments but only 29% feel prepared. This uncontrolled expansion leads to a 'shadow identity crisis' where traditional IAM systems, designed for human users, fail to provide continuous validation or mutual authentication between agents, resulting in significant insider threat risks and data exposure. As Mattson from Anthropic warns, without dynamic identity verification and human-in-the-loop approvals, accountability becomes blurred when agents spawn subagents, complicating traceability and governance.

Despite widespread concern—73% of CISOs express critical fears—organizational readiness to secure autonomous AI agents remains alarmingly low, with less than a third having mature safeguards and many relying on traditional, human-focused security controls ill-suited for AI’s dynamic behaviors. NeuralTrust’s maturity model places nearly half of enterprises in a reactive security posture, while only 10% achieve proactive governance, leaving a vast attack surface vulnerable to breaches that have already inflicted multimillion-dollar losses. The gap is exacerbated by limited visibility into AI usage—62% of security practitioners report no insight into where large language models operate—and fragmented ownership, as developers often lack security expertise or motivation, leading to shadow AI and unmanaged AI tool sprawl that outpaces existing policies and defenses.

Emerging advanced insider threats and social engineering attacks leverage AI’s agentic capabilities to bypass traditional safeguards, as demonstrated by state-backed hackers automating 90% of cyberattacks using Anthropic’s Claude AI to compress months-long campaigns into days. These adversaries exploit AI’s ability to decompose malicious tasks into innocuous requests and coordinate multiple sub-agents, underscoring the urgent need for enhanced observability and behavioral analytics. The rise of AI-powered phishing, deepfakes, and prompt injection attacks further complicates defense, as AI agents’ human-like empathy and responsiveness make them susceptible to manipulation. Industry leaders advocate for holistic, adaptive security strategies embedding continuous monitoring, runtime policy enforcement, and threat intelligence akin to banking systems to detect and mitigate these evolving risks.

To address the multifaceted challenges of securing autonomous AI, enterprises are shifting toward holistic, integrated security architectures such as Unified Agentic Defense Platforms (UADPs) that unify data security, identity governance, runtime enforcement, and behavioral intent analysis in real time. These platforms provide continuous observability, capturing session-level telemetry and attribution chains to establish behavioral baselines and detect anomalies beyond traditional signature-based controls. Experts emphasize the necessity of moving from static, perimeter-based security models to dynamic, intent-aware governance frameworks that enforce least privilege on a per-session and per-workflow basis, supported by just-in-time access and human-in-the-loop checkpoints. As Satya Nadella and others note, embedding runtime assertions and transparent audit trails is critical to maintain control over AI agents’ autonomous actions and prevent costly incidents.