Shadow AI pushes boards toward control measures

The Hacker News ↗

The gist

Shadow AI is exploding across enterprises, blindsiding security teams and forcing boardrooms to race for control as attack surfaces outgrow traditional IT defenses.

What to know

  • By late 2025, 62% of security practitioners admitted they had no idea where large language models were running, fueling a shadow AI crisis that now eclipses classic shadow IT headaches.
  • Traditional security models are failing against autonomous AI agents, prompting a shift to zero trust architectures, real-time session verification, and continuous permission checks to fight emerging threats like credential hijacking.
  • Defense and enterprise leaders are demanding sovereign, on-premises AI and stronger human oversight—balancing rapid innovation with new governance frameworks to avoid vendor lock-in and rogue machine decisions.

Blind Spots in AI Security

AI-native apps and autonomous agents are outpacing security teams, creating sprawling, invisible attack surfaces that legacy tools can’t monitor or control.

By late 2025, enterprises were already grappling with a profound AI visibility crisis, with 62% of security practitioners reporting no insight into where large language models were deployed, creating a fertile ground for shadow AI to emerge as the new blind spot surpassing even shadow IT risks. This rapid AI sprawl, fueled by the proliferation of AI-native applications viewed by 82% of respondents as prime targets for cybercriminals, expanded attack surfaces beyond traditional IT, while only a minority of developers engaged security teams proactively, exacerbating governance challenges.

The autonomous nature of AI agents, which often operate with broad, long-lived permissions and can execute commands on host machines, has introduced unprecedented security blind spots that traditional tools and static security models cannot address. As highlighted by Reco’s research showing that 80% of AI tools run without IT oversight and many expose unauthenticated network endpoints, enterprises face an urgent need to rethink security architectures to include continuous runtime monitoring and strict permission controls to mitigate risks such as credential theft, agent hijacking, and data exfiltration.

The rapid pace of AI adoption has outstripped the development of security policies and expertise, leaving organizations vulnerable to emerging threats like session hijacking and universal AI jailbreaks that bypass model restrictions regardless of vendor. This gap is compounded by fragmented security stacks and tool sprawl, with 75% of CISOs relying on legacy controls ill-suited for adaptive AI systems, and half of organizations experiencing AI-related incidents despite deploying AI security measures, underscoring a critical need for unified platforms and AI-driven security solutions.

Enterprise leaders express growing fear and frustration over losing control and visibility of AI agent activities within their environments, as these non-human actors rapidly execute complex tasks with legitimate access but minimal oversight. This proliferation of shadow AI agents, often introduced through unsanctioned means like OAuth consent screens or browser extensions, complicates accountability and governance, making traditional human-centric security models obsolete and demanding real-time action controls and comprehensive visibility to manage the expanding, unpredictable AI attack surface.

Sources

Boardroom Battles Over AI Governance

Shadow AI and unchecked vendor reliance are forcing executives to overhaul governance, as hidden AI deployments threaten compliance and business continuity.

As enterprises rapidly adopt AI under board and CEO mandates, governance frameworks must evolve to address critical risks stemming from vendor concentration and shadow AI proliferation. Deepen Desai highlights that reliance on a few dominant AI providers creates vulnerabilities not only in security but also in business continuity, especially when embedded AI usage within supply chains goes unnoticed. This challenge is compounded by the widespread presence of shadow AI—unapproved or unknown AI tools operating across departments—posing significant compliance and oversight difficulties that demand elevated executive attention.

The emergence of autonomous AI agents with broad, multi-system access has intensified the need for integrated AI governance frameworks that span development through deployment, combining security, compliance, and operational agility. Partnerships like Check Point and ControlPlane exemplify this approach by delivering AI-native threat prevention alongside cloud-native DevSecOps, enabling regulated industries to innovate without sacrificing compliance. However, traditional security controls remain insufficient, prompting calls from cybersecurity leaders like Jeetu Patel and Emil Michael to embed governance deeply into AI workflows and elevate AI security to the boardroom to manage risks such as prompt injection, data leakage, and model misuse.

Data sovereignty and regulatory compliance have become paramount governance concerns, especially in defense and national security contexts where control over AI models and data provenance directly impact operational sovereignty. Initiatives like Defence Australia’s One Defence Data program and Naver Cloud’s Data Fabric architecture illustrate efforts to maintain data within secure, federated environments while enabling AI-driven insights. Yet many enterprises remain unaware of contractual pitfalls allowing vendors to use anonymized data for AI training, exposing organizations to unintended competitive risks. Board-level oversight is crucial to enforce distinct governance models at enterprise, project, and user levels, ensuring proprietary intelligence is protected and compliance obligations are met.

Despite widespread AI adoption, governance maturity lags significantly, with studies revealing that 80% of AI tools operate without IT oversight and half of organizations experiencing AI-related incidents even with security controls in place. This gap is driven by fragmented security stacks, skill shortages, and reliance on legacy controls ill-suited for AI’s unique risks. Experts emphasize the necessity of board-level accountability to drive integrated governance frameworks that manage shadow AI, enforce least privilege access, and maintain continuous monitoring. As Arvin Nithra Kashayap and others note, the velocity and autonomy of AI agents require human-in-the-loop models and strategic executive engagement to prevent costly breaches and ensure enterprise resilience.

Sources
Insights by KPIT Brief New ZealandThe Tech TrekBFGlobeNewswire - Industry News on TechnologyThe Hacker News

Zero Trust for Autonomous Agents

Static security is dead—continuous, dynamic verification and least-privilege autonomy are now essential as AI agents operate with unprecedented freedom and risk.

Traditional static security models that rely on one-time authentication and broad, long-lived credentials have proven inadequate for securing autonomous AI agents, as attackers increasingly hijack session credentials and exploit delegated permissions. The emerging concept of 'session as the new perimeter' underscores the necessity of continuous, dynamic verification throughout AI agent sessions to prevent replay attacks and agent hijacking, as highlighted by the 2025 SalesLoft breach and subsequent analyses emphasizing zero trust protocols and granular session checks.

By early 2026, industry leaders and security researchers recognized that legacy IAM and perimeter-based controls fall short for agentic AI, prompting a shift toward least permissive autonomy and trust engineering frameworks. Incidents like the OpenClaw compromise exposed critical vulnerabilities from publicly exposed credentials, driving adoption of Policy-as-Code tools like Cedar and integrated AI-native threat prevention platforms from Check Point and ControlPlane. These solutions emphasize least privilege access, continuous monitoring, and human-in-the-loop approvals to secure AI agents operating across cloud, SaaS, and endpoint environments at scale.

The rapid proliferation and autonomous capabilities of AI agents have shattered traditional security assumptions, demanding evolved zero trust architectures that treat AI agents as non-human identities with dynamic, ephemeral permissions. Experts like Adid Shamir and Dimitri Sirota stress expanding least privilege into 'least agency' and embedding trust engineering from the outset to manage AI’s improvisational behavior and compressed vulnerability-to-exploit timelines—what some call the 'Vulnpocalypse.' This requires defense-in-depth across data, model, and agent layers, continuous behavioral monitoring, and containment strategies to mitigate risks posed by multi-agent coordination and AI-driven supply chain attacks.

Despite the critical role of human oversight in governing AI agents, security architectures must move beyond reactive, human-in-the-loop models toward proactive, action-driven frameworks that integrate continuous verification, red teaming, and runtime guardrails. Leading organizations like Google, Cisco, and Datadog advocate for hardened authorization boundaries, scoped credentials, and real-time anomaly detection to prevent unauthorized or malicious AI actions. This approach includes purpose-built tools that enforce least privilege access, input validation, and emergency revocation, ensuring AI agents operate within tightly controlled environments without impeding innovation or developer productivity.

Sources
HNIBM TechnologyResilient CyberBusiness WirePR Newswire - Business TechnologySuper Data Science: ML & AI Podcast with Jon Krohn

AI Supply Chains: The New Risk Frontier

Unseen AI integrations across supply chains are creating systemic vulnerabilities, as organizations struggle to map, secure, and govern hidden AI dependencies.

Unseen AI integrations across supply chains are creating systemic vulnerabilities, as organizations struggle to map, secure, and govern hidden AI dependencies.

Sovereign AI: Defense’s New Imperative

National security leaders are demanding on-premises, customizable AI to avoid vendor lock-in and prevent external ‘kill switches’ from undermining military operations.

By early 2026, defense organizations recognized that the rapid adoption of AI in military contexts is a high-stakes balancing act between speed and caution. While the military acknowledges that delaying AI integration risks operational disadvantage, there is strong skepticism about entrusting sensitive military data and weapon systems to certain AI vendors, notably Elon Musk’s xAI, due to concerns over reliability and alignment. This tension underscores the imperative for sovereign AI infrastructure that can uphold rigorous standards of precision and ethical responsibility without succumbing to reckless deployment or vendor lock-in.

Sovereign AI deployments are increasingly defined by enterprises and defense agencies demanding full control over their data and AI models, often opting for on-premises or virtual private cloud environments to avoid the risks of data movement and external exposure. Companies like Mistral AI emphasize customization through fine-tuning with proprietary data, enabling models to adapt to specific languages and workflows while maintaining strict privacy. This approach, supported by dedicated teams of AI engineers and applied scientists, addresses operational security and governance challenges inherent in defense and sovereign contexts, mitigating risks such as vendor lock-in and data leakage.

The defense ecosystem faces systemic risks from heavy reliance on proprietary AI vendors, which can create dangerous vendor lock-in and supply chain vulnerabilities, as highlighted by the Anthropic controversy where refusal to remove usage restrictions led to government classification of supply chain risk. Such dependencies risk operational failures if AI models are deactivated mid-mission due to contractual or geopolitical constraints, effectively granting external entities a 'kill switch' over critical military capabilities. This dynamic has propelled nations like the U.S., France, and South Korea to invest in sovereign AI infrastructure, emphasizing physical security, local hosting, and closed networks to maintain operational autonomy and prevent foreign control.

Data sovereignty has emerged as a foundational pillar in securing AI across enterprise and defense ecosystems, requiring zero-copy architectures that keep data in place and enforce strict access controls to prevent unauthorized cross-departmental data exposure. Distributed AI techniques, such as swarm learning, enable knowledge sharing without data sharing, addressing stringent regulatory and privacy demands. Defense agencies like Defence Australia prioritize federated data architectures that connect disparate sources without centralization, ensuring real-time, trusted information delivery while preserving classification boundaries. This comprehensive approach integrates hardware-enforced secure fabrics and proactive governance to safeguard AI workflows from prompt injection attacks and unauthorized data movement, recognizing that the true AI security risk lies in data access rather than the models themselves.

Sources
CBBFImagination in ActionAxiosAxiosThe Hacker News

Humans in the Loop—Or Out?

As AI agents gain real-world autonomy, maintaining human oversight and local trust networks is the last line of defense against unaccountable machine decisions.

As AI increasingly transitions from generating digital insights to executing physical actions, maintaining human agency and oversight becomes paramount to ensure trust, security, and alignment with human intent. The strategic imperative to avoid outsourcing critical decisions—especially in warfare—to autonomous systems is underscored by concerns that detaching humans from lethal realities risks making warfare a slave to machine autonomy. Models like human-in-the-loop and human-on-the-loop are essential frameworks that balance AI assistance without supplanting human control, ensuring accountability remains firmly with humans rather than AI agents.

The rapid pace of AI advancement, exemplified by ChatGPT's leap from version 5 to 5.2 in under nine months, compresses traditional risk assessment timelines and challenges existing governance frameworks. This acceleration demands a shift from attempting to prove AI outputs individually toward defining and enforcing system boundaries through harness engineering, continuous authorization, and zero trust principles. Experts like Will Pearce and Victoria Westeroff emphasize that AI agents require live permission checks and risk-based human oversight focused on high-stakes decisions, enabling scalable innovation while managing inherent non-deterministic behaviors.

AI adoption is increasingly embedded within existing trust networks—such as messaging platforms, creator workflows, and education systems—rather than through standalone applications, highlighting that trust remains a fundamentally local and relational phenomenon. This trend drives national and enterprise-level requirements for sovereign compute, local model hosting, and regulatory oversight to build secure, resilient AI infrastructures. As AI agents act as extensions of trusted individuals, governance must evolve to address distribution and trust challenges alongside technological innovation, ensuring that pluralistic ecosystems can coexist without centralized control dominating information consumption.

Right-sizing AI adoption within defense and enterprise ecosystems involves implementing robust guardrails, cybersecurity, and adaptable governance processes that respect fluid and evolving data environments. Leaders stress the importance of a secure digital foundation with comprehensive asset and data visibility to transition from reactive to proactive AI management. Avoiding extremes—such as outright AI bans or wholesale enterprise transformations without readiness—ensures AI tools deliver optimal value while maintaining security and governance integrity, a balance critical to sustaining innovation without compromising resilience.

Sources
Fed Gov TodayCognitive Revolution "How AI Changes Everything"CyberWire DailyTo The Point - CybersecurityTEa16z

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.