Identity-first security, runtime protections, and governance for AI agents

Securing Agentic AI

The accelerating integration of autonomous, agentic AI systems into enterprise, cloud, edge, and operational technology (OT) environments continues to drive a pivotal evolution in cybersecurity strategies. The urgency to adopt identity-first security architectures, adaptive runtime protections, and unified governance frameworks tailored for AI agents has never been more pronounced. Recent developments further reinforce and expand these foundational pillars, introducing novel vendor collaborations, comprehensive market insights, and practical demonstrations that exemplify holistic AI agent security in real-world workflows.

Identity-First Security: Expanding the Cryptographic Foundation of AI Agents

AI agents now possess cryptographic identities that operate as first-class security principals, requiring robust protections against a spectrum of emerging threats, particularly quantum and hardware-based attacks. Building on prior advances, new initiatives highlight:

Industry-Wide Acceleration of Post-Quantum Cryptography (PQC):
Adoption of lattice-based PQC algorithms continues to gain momentum, particularly for securing AI agent keys and tokens across hybrid cloud and edge deployments. Vendors like Palo Alto Networks are embedding PQC as a core feature within their agent security stacks, showcasing a proactive, future-proof approach that anticipates the maturation of quantum computing threats.
Hardware-Rooted Secrets and Lifecycle Management:
The standardization of using TPMs, secure enclaves, and HSMs for storing AI agent credentials is complemented by advanced techniques such as ephemeral credential issuance and continuous rotation. Palo Alto Networks’ acquisitions of Koi Security and CyberArk underscore a broader vendor trend toward tightly integrated cryptographic identity and privileged access management systems for AI agents.
Kernel-Level Identity Anchors and Policy Binding:
Embedding cryptographic identities directly into the OS kernel or AI runtime creates tamper-resistant trust anchors that significantly raise the bar against impersonation and privilege escalation. This approach, while powerful, requires sophisticated kernel hardening to mitigate risks of centralized failure points.
Dynamic Identity-Linked Access Control:
Tools like Tailscale’s Aperture, now in open alpha, exemplify how cryptographic agent identities dynamically enforce zero-trust access policies across cloud, edge, and hybrid environments. This fine-grained control reduces insider threat surfaces and aligns operational security with cryptographic identity management.
Vendor Innovations and Ecosystem Expansion:
- Vast Data’s AI Operating System now includes a global control plane paired with a zero-trust agent framework leveraging NVIDIA hardware, enabling unified cryptographic enforcement across AI workloads.
- Netskope NewEdge AI Fast Path optimizes network performance for AI traffic while embedding continuous security monitoring.
- Zenarmor’s SASE Channel Partner Program accelerates AI-aware Secure Access Service Edge deployments, integrating identity, telemetry, and network controls for consistent zero-trust enforcement.
- Hypori and Menlo Security Collaboration: A newly released video demonstrates end-to-end workday security patterns that harmonize identity-first security, browser isolation, and runtime protections—delivering a practical blueprint for securing agentic AI workflows in enterprise contexts. This joint effort underscores the increasing importance of isolated browsing environments in minimizing attack surfaces for AI-driven interactions.

Adaptive Runtime Protections: Real-Time Defense in an Evolving Threat Landscape

Static and perimeter-based defenses cannot keep pace with the dynamic, autonomous behaviors of agentic AI. New runtime protection mechanisms emphasize continuous, context-aware monitoring and response:

RAG-Aware Inference Controls Become Mainstream:
Integrating Retrieval-Augmented Generation awareness into runtime security allows organizations to detect and prevent retrieval poisoning and inadvertent data leakage during AI inference, a critical advancement in protecting sensitive corpora accessed or generated by AI agents.
Sophisticated Behavioral Monitoring and Runtime DLP:
Real-time AI behavioral analytics identify anomalous API calls and suspicious data movements. Endpoint DLP solutions now effectively block unauthorized uploads of corporate data to generative AI platforms. BlackFog champions this shift toward anti-data-exfiltration as a new cybersecurity baseline, embedding continuous data flow controls within AI-assisted workflows.
Microsegmentation and Network Isolation:
Projects like Claws demonstrate how iptables and security groups can enforce strict network segmentation of AI agents, limiting lateral movement and reducing breach impact, especially critical in complex hybrid cloud-edge-OT environments.
Dynamic Risk Scoring and Adaptive Policy Enforcement:
Enterprise browsers including Microsoft Edge and dME incorporate AI-driven behavioral signals with human user context, enabling real-time adaptive risk scoring that adjusts access policies dynamically, enhancing runtime protection tied directly to evolving threat intelligence.
AI-Powered Detection and Automated Incident Response:
Platforms like Cato Networks’ AI Security for Applications leverage behavioral analytics to detect runtime anomalies and unauthorized model changes. With attacker dwell times shrinking to an average of 72 minutes (per Unit42’s 2026 report), automated containment and forensic triage augmented by human oversight are now imperative.
Regulatory Push for Runtime Security:
The Cybersecurity and Infrastructure Security Agency’s (CISA) Emergency Directive 26-03 mandates urgent mitigation of vulnerabilities in Cisco SD-WAN systems—highlighting the critical need to secure edge and OT network components that interface with AI agents. Complementing this, the recently released Frost Radar™ Managed SD-WAN (2025) report offers an in-depth market analysis detailing vendor capabilities, secure deployment strategies, and innovation trends for managed SD-WAN in hybrid cloud and edge environments. This report provides valuable insights for organizations aligning their SD-WAN infrastructures with CISA’s stringent security directives.

Unified Governance and Telemetry: Ensuring Transparency and Accountability Across AI Ecosystems

As AI agents integrate deeper into mission-critical operations, governance frameworks must deliver comprehensive visibility, enforce compliance, and facilitate accountability:

Holistic Telemetry Aggregation:
Cutting-edge solutions unify telemetry streams from human users, AI agents, devices, and network identities into centralized platforms. This end-to-end visibility spans the entire AI lifecycle—from model training and deployment to retirement—supporting forensic investigations, real-time monitoring, and compliance audits.
Identity-Linked Governance with Immutable Auditing:
Frameworks that bind cryptographic agent identities to access policies and immutable audit logs reduce insider threats and ensure traceability. Tailscale’s Aperture leads in enforcing such dynamic policies across hybrid infrastructures.
Cross-Cloud and OT Policy Consistency:
Enterprises demand seamless zero-trust enforcement across AWS, Azure, Google Cloud, and increasingly integrated OT environments. NVIDIA’s partnerships with cybersecurity vendors aim to extend zero-trust protections into Industrial Control Systems (ICS), addressing the unique complexities introduced by agentic AI in OT.
Sectoral and Regulatory Alignment:
- The U.S. Department of the Treasury’s 2026 AI Guardrails set stringent standards for financial institutions, emphasizing transparency, risk assessment, and incident reporting—aligned with the NIST AI Risk Management Framework (AI RMF 2024–2025).
- Healthcare stakeholders, led by the American Hospital Association, advocate for zero-trust architectures explicitly incorporating AI agents to safeguard patient confidentiality and regulatory compliance.
- Sovereign AI initiatives, exemplified by Lanarkshire’s regional efforts, harmonize governance frameworks with data sovereignty laws to enable innovation within legal boundaries.
- The CISA Binding Operational Directive (BOD) 26-02 mandates comprehensive lifecycle management for edge devices, closing critical governance gaps at network perimeters.
Collaborative Industry Initiatives for Integrated Governance:
The joint Google and Microsoft WebMCP initiative embeds cryptographic provenance and strict data access policies into AI browsing workflows, mitigating risks from malicious browser extensions and “shadow AI” agents. This collaboration epitomizes the industry-wide drive toward integrated governance models uniting identity, data, and runtime behavior.

Implications and Outlook: Toward a Resilient, Identity-First Security Paradigm for AI Agents

The evolving AI threat landscape demands a multi-layered, identity-first security architecture anchored in cryptographic trust, hardware-backed protections, and adaptive runtime defenses. Recent developments emphasize:

Cryptographic Identities as Security Foundations:
AI agents require tamper-resistant, hardware-backed digital identities governed through robust lifecycle management—covering issuance, rotation, recovery, and revocation.
Proactive Runtime Protections:
Real-time behavioral analytics, RAG-aware inference controls, runtime DLP, microsegmentation, and adaptive risk scoring combine to thwart sophisticated adversarial tactics before they can cause damage.
Unified Telemetry and Governance for Accountability:
Single-pane-of-glass visibility spanning human and AI entities enables comprehensive monitoring, forensic readiness, and regulatory compliance in complex hybrid environments.
Vendor Ecosystem and Regulatory Synergy:
Integration of innovative solutions from Palo Alto Networks, Netskope, Vast Data, Tailscale, Zenarmor, Hypori, and Menlo Security, alongside compliance with government mandates from Treasury, NIST, and CISA, equips organizations to deploy secure, scalable AI agent ecosystems.
Emerging Standard of Anti-Data Exfiltration:
The shift beyond perimeter defense to embedded, continuous data flow monitoring ensures that sensitive information remains protected throughout AI-augmented workflows.
Practical Workday Security Patterns:
The Hypori + Menlo Security video demonstration offers a tangible example of securing AI-driven workflows end-to-end, blending identity-first principles, browser isolation, and runtime protections—providing a replicable model for enterprises.
SD-WAN Security and Market Trends:
CISA’s emergency directives and the Frost Radar™ Managed SD-WAN report together highlight the critical role of secure SD-WAN deployments in hybrid cloud and edge environments where AI agents operate, emphasizing the convergence of connectivity, security, and governance.

Organizations that embrace these integrated, identity-first security frameworks will be positioned not only to mitigate the unique risks posed by agentic AI but also to unlock its transformative potential—driving innovation while safeguarding operational resilience in an increasingly adversarial ecosystem.

References for Further Exploration

Tailscale launches Aperture in open alpha for identity-linked governance of AI tools and agents
Palo Alto to acquire Israeli startup Koi for agentic AI security
Vast Data expands AI Operating System with global control plane, zero-trust agent framework and deeper Nvidia integration
Netskope NewEdge AI Fast Path reduces latency for enterprise AI workloads
Zenarmor Creates Architecture-Driven SASE Channel Partner Program
Hypori + Menlo Security: Securing the Entire Workday (YouTube video)
Frost Radar™: Managed SD-WAN in North America, 2025 (PDF report)
Cato AI Security for Applications
Treasury releases new guidelines for responsible use of artificial intelligence in finance
NIST's AI Risk Management Framework in 2024-2025
Claws are now a new layer on top of LLM agents | Hacker News
Beyond the Perimeter: Anti Data Exfiltration is the New Cybersecurity Standard
Google & Microsoft Want To Fix AI Browsing (With WebMCP)
CISA BOD 26-02 Signals a New Era of Edge Device Lifecycle Accountability
Zscaler Data Security Services Explained — Zero Trust for Your Data
The ABCs of Securing Agentic AI: Protecting Agents, Browsers, and Co-Pilots
CISA Issues Emergency Directive to Secure Cisco SD-WAN Systems

The continued convergence of cryptographic identity management, runtime protections, and unified governance forms the cornerstone of resilient AI agent security—empowering enterprises to innovate responsibly and securely in the AI era.

Sources (66)