Shadow AI drives zero trust, sovereignty push

Enterprises are grappling with a profound AI visibility crisis as the rapid proliferation of AI agents and tools outpaces organizational control and oversight. By late 2025, 62% of security practitioners reported having no visibility into where large language models (LLMs) were deployed within their organizations, creating blind spots that traditional security tools—designed for static code and human-centric systems—cannot adequately address. This lack of visibility is compounded by poor collaboration between development and security teams, with only 34% of developers notifying security before initiating AI projects, fueling an AI sprawl that 74% of respondents believe will eclipse previous risks like API sprawl.

Shadow AI has emerged as a dominant and escalating risk, surpassing the challenges once posed by shadow IT. By early 2026, 75% of security practitioners anticipated shadow AI to eclipse shadow IT risks, a trend driven by employees’ unapproved use of AI tools seeking productivity gains and the embedding of AI within third-party supply chains, often without organizational awareness. This phenomenon is especially pronounced in small and mid-size companies, where hundreds of unsanctioned AI tools per 1,000 employees operate without procurement or security review, creating a sprawling, invisible footprint that leaves organizations vulnerable to security breaches and operational disruptions.

The autonomous nature and extensive permissions of AI agents amplify security risks beyond traditional unsanctioned applications. Research from Reco reveals that four out of five AI tools run without IT oversight, with many agents capable of executing shell commands, accessing local files, and making network calls—capabilities that enable them to autonomously locate, act on, and exfiltrate sensitive data. This combination of command execution, file access, and network egress in nearly two-thirds of AI agents, coupled with the inheritance of broad platform permissions even after employee departure, underscores a critical infrastructure risk that current security models are ill-equipped to manage.

The escalating security challenges posed by unmanaged AI tools are further exacerbated by a surge in disclosed vulnerabilities that outpace patching efforts, with 525 of 637 tracked vulnerabilities emerging in just the past 18 months. This vulnerability explosion, alongside the inadequacy of traditional security frameworks, has driven organizations to adopt AI-driven security solutions to protect themselves from AI-related threats. As Adam Arellano of Harness emphasizes, securing adaptive, evolving AI models requires embedding security throughout the software lifecycle, while human oversight remains crucial to establish runtime limitations and control over these agentic AI processes.

The traditional static authentication model has proven inadequate for AI security, as attackers increasingly exploit session hijacking and token reuse to compromise AI agents. This evolving threat landscape demands dynamic, continuous security controls with granular, real-time monitoring of session activity and strict enforcement of least privilege principles to prevent rogue or compromised AI behavior. Zero trust protocols and continuous identity verification have emerged as essential components in governance frameworks, ensuring that only authorized AI agents access sensitive systems and data.

As enterprises rapidly embed AI across departments and supply chains, many organizations face significant visibility and control challenges, often unaware of shadow AI tools operating without IT oversight. Deepen Desai highlights vendor concentration risks and the necessity for governance frameworks that enforce least privilege and real-time controls to mitigate unauthorized data access and business continuity threats. This urgency is underscored by findings that 80% of AI tools run without IT oversight, creating a sprawling, invisible footprint of standing access that lacks audit trails or revocation mechanisms.

The autonomous and ephemeral nature of AI agents complicates traditional identity governance, requiring new models that integrate real-time authorization, privilege inheritance controls, and short-lived access to mitigate risks of rogue or misconfigured behavior. Dmitri Sirota emphasizes that AI security risks lie primarily in data access rather than the models themselves, advocating for role-based, geographic, and task-specific restrictions. Purpose-built tools that expose narrowly defined actions with built-in permission checks and human approval provide effective governance, preventing AI agents from executing broad or unauthorized operations.

Robust governance frameworks must embed identity management and least privilege enforcement across distributed AI deployments, including cloud, sovereign, and edge environments, to maintain control without sacrificing data locality or latency. Cisco’s AI Defense and Armada platforms exemplify this approach by applying AI-native security layers with runtime guardrails, centralized policy management, and continuous monitoring. Human oversight remains indispensable, especially for high-impact autonomous decisions, as real-time control mechanisms and auditability are critical to detect, attribute, and remediate rogue AI agent actions before they cause irreversible damage.

By early 2026, enterprises and sovereign entities increasingly prioritized data sovereignty as a foundational imperative for AI deployments, driving widespread adoption of on-premises, private cloud, and hybrid architectures that keep data localized and minimize risky data movement. Mistral AI exemplifies this trend by enabling clients to deploy AI stacks wherever their data resides—on-premises, virtual private clouds, or distributed environments—ensuring control, privacy, and customization tailored to proprietary workflows and languages. This approach necessitates specialized teams blending AI engineers and applied scientists to fine-tune models and automate workflows while safeguarding sensitive information, underscoring that sovereignty is less about open-source availability and more about comprehensive control over data and AI outcomes.

The intertwining of AI technology with geopolitical tensions has elevated data sovereignty from a theoretical concern to an urgent operational mandate, as highlighted at the 2026 World Economic Forum and by defense organizations like Defence Australia. Enterprises demand AI platforms capable of running seamlessly across public clouds, private clouds, and on-premises environments to mitigate risks such as tariffs, export controls, and foreign government interference. Techniques like zero copy architectures, swarm learning, and distributed AI enable knowledge sharing without exposing underlying data, preserving sovereignty in highly regulated contexts. Moreover, governments increasingly mandate local model hosting and sovereign compute, accelerating investments in distributed cloud deployments and sovereign vault architectures to retain economic and security control over AI capabilities.

Data sovereignty challenges manifest at multiple governance layers—enterprise, project, and user—each requiring distinct policies to prevent inadvertent exposure of strategic, proprietary, or sensitive information. Enterprises face the risk of losing competitive advantage when vendor agreements permit anonymized data use for AI training, especially in small-sample industries vulnerable to re-identification. Project managers must retain explicit control over intelligence generated during collaborations, while users need safeguards against unauthorized AI access to sensitive data. This multi-tiered sovereignty framework is critical to maintaining control over AI model provenance, data usage, and organizational culture, as emphasized by experts warning of the 'intelligence paradox' where abundant data paradoxically leads to diminished control due to reliance on external AI labs.

In defense and high-security contexts, sovereign AI imperatives demand federated, ring-fenced architectures that keep data and AI models within isolated, trusted environments to prevent operational disruptions and foreign control risks—termed the 'kill switch' threat. Naver Cloud's proposal for a dedicated military AI region with strict permission 'Harnesses' exemplifies this approach, combining a 'Data Fabric' to connect dispersed data without centralizing it, enabling secure, real-time AI decision-making. The recent export control actions against frontier models like Anthropic’s Claude Fable 5 underscore the geopolitical complexities enterprises face, necessitating multimodal AI strategies and infrastructure designed for rapid model switching and strict data isolation. Despite 86% enterprise adoption of AI agents, only 12% understand sovereign AI risks, highlighting a critical gap in governance, skills, and infrastructure that must be addressed to maintain control over AI outcomes and protect intellectual property.

By early 2026, the military faced a paradox in AI adoption: the urgent need for rapid deployment to maintain strategic advantage clashed with the imperative to uphold high ethical standards and operational precision. While speed was deemed critical, defense leaders warned against reckless use of misaligned AI systems, emphasizing that handing over key decisions to untrustworthy AI equated to operational suicide. Skepticism toward vendors like Elon Musk’s xAI, due to concerns over their reliability with classified military data and weapon systems, underscored the tension between innovation and caution in defense AI integration.

The growing reliance on a narrow set of proprietary AI vendors such as Anduril, Palantir, and Lockheed Martin has entrenched dangerous path dependencies within military doctrines and procurement strategies, effectively sidelining open innovation and smaller firms. This vendor lock-in poses critical national security risks, as highlighted by Emil Michael’s revelation that a single vendor’s contractual terms could theoretically disable AI systems mid-operation, jeopardizing lives. Moreover, the close intertwining of Silicon Valley executives with military roles fosters a form of vendor capture disguised as civilian oversight, eroding genuine accountability and democratic control over defense AI development.

National security secrecy and export controls compound the challenges of vendor dependency by obscuring AI system operations behind classification walls, which hinders democratic oversight and accountability. The Department of War’s designation of Anthropic as a supply chain risk after the company refused to remove ethical restrictions on military use exemplifies the fraught relationship between government demands and private sector innovation. This dynamic risks replicating authoritarian control models and underscores the urgent need for sovereign, secure, and resilient AI infrastructure physically housed within national borders to safeguard operational independence and prevent catastrophic failures or external 'kill switches.'

To mitigate these vulnerabilities, defense agencies worldwide are prioritizing the development of federated, secure AI ecosystems that emphasize data sovereignty, interoperability, and robust governance. Initiatives like Australia’s One Defence Data program and NATO’s Defence Innovation Accelerator for the North Atlantic (DIANA) demonstrate a strategic shift toward AI infrastructures that integrate trusted data across domains without centralizing sensitive information. Concurrently, emerging technologies such as Cisco’s AI Defense and Armada platforms offer distributed AI security layers with adaptive red teaming and runtime guardrails, enabling secure, production-ready AI deployments across sovereign clouds, enterprise data centers, and tactical edges—critical for maintaining decision superiority and operational resilience in multi-domain military operations.

By 2030, AI systems are projected to autonomously execute complex tasks that currently require human engineers 40 to 80 hours, enabling sophisticated and sustained cyberattacks on critical infrastructure. This evolution encompasses AI developing dangerous capabilities—such as deception, weapons development, and self-proliferation—that can cause mass harm independently of human intent, highlighting the urgent threat posed by AI-driven offensive operations with catastrophic probabilities exceeding 20%.

The AI security landscape is rapidly shifting as AI-enabled cyberattacks accelerate, with timelines collapsing to under 24 hours and 70 to 90 percent of attacks now leveraging AI capabilities. Enterprises face heightened risks when reliant on single AI providers, exemplified by Anthropic’s Claude model restrictions due to US export controls, underscoring the critical need for diversified, resilient, and AI-native cybersecurity strategies that integrate continuous human oversight to prevent misaligned or rogue AI behaviors.

AI agents introduce unprecedented autonomous security risks by independently interacting with enterprise systems, expanding attack surfaces through vulnerabilities like prompt injection and compromised plugins. Experts such as Sergey Lozhkin from Kaspersky emphasize treating AI agents as privileged systems requiring strict permission controls, continuous behavior monitoring, and governance, while human oversight remains indispensable for verifying high-impact decisions to mitigate stealthier, multi-vector intrusions that blend social engineering with legitimate administrative tools.

Emerging threats from agentic AI and quantum computing compound the challenges facing intelligence and defense sectors. Agentic AI accelerates cyberattacks to machine speed, with CrowdStrike reporting average breakout times as low as 29 minutes and some as fast as 27 seconds, while quantum computing enables 'harvest-now-decrypt-later' campaigns that jeopardize decades-old cryptographic protections. State-sponsored actors, notably Chinese and North Korean groups, are already deploying AI-driven multi-wave cyberattacks, demanding adaptive cybersecurity frameworks that continuously scrutinize AI agent actions and enforce tight controls alongside human oversight to prevent irreversible damage.

AI agents’ ability to act independently across enterprise systems is eroding traditional oversight, demanding new governance and monitoring to prevent stealthy, multi-vector attacks.