Key takeaways
- Securing agentic AI in government requires treating non-human AI agents as distinct identities governed by Zero Trust architecture.
- Government agencies must implement continuous identity discovery, least-privilege access controls, real-time runtime monitoring, and immediate lifecycle revocation mechanisms (kill switches) to prevent unauthorized access to sensitive public-sector data.
As we mark Cybersecurity Awareness Month, I find myself reflecting on the common thread connecting my time managing federal cybersecurity at the Department of Justice to my work today at Okta supporting enterprises and government agencies.
Throughout my career, whether safeguarding sensitive case data at the Executive Office for United States Attorneys or helping Fortune 500 companies modernize their tech stacks, one truth has remained constant: You cannot protect what you cannot identify.
Today, as artificial intelligence reshapes both the public and private sectors, that lesson is more critical than ever. AI offers extraordinary power: automating citizen services, streamlining complex supply chains, and processing vast data sets in seconds. But without a strong security foundation, autonomous tools and agents can quickly turn into unmonitored backdoors to sensitive data.
Why is agentic AI security fundamentally an identity problem?
Think of AI as an incredibly fast, highly capable assistant. To do its job, this assistant needs access to your files, databases, and internal communications.
If you wouldn't hand the keys to your evidence lockers or filing cabinets to an unverified contractor, you shouldn't let an unverified user or automated bot interact with your AI system. AI security is ultimately an identity problem.
A blueprint for the secure agentic enterprise
To tackle this, Okta co-founded the Blueprint Alliance—an open industry initiative establishing a standardized reference architecture for securing the agentic enterprise. At its core, the Alliance organizes agentic security around four fundamental questions every organization must be able to answer:
- Where are my agents?
- What can they do?
- What are they doing?
- How do I respond?
Securing AI through identity: The four essential pillars for securing agentic AI in government
1. Identity discovery and verification: "Where are my agents?"
- The challenge: You cannot protect or govern what you cannot see. Shadow AI tools and rogue bots must be brought into the open and registered with clear operational ownership.
- The solution: Before any employee, contractor, or automated system can send prompts (or before an agent acts autonomously) its identity must be discovered, cataloged, and verified beyond doubt.
- Okta implementation: Okta acts as the digital gatekeeper. Through phishing-resistant multi-factor authentication (MFA) and continuous identity evaluation, Okta ensures only verified human users and authorized software services can access your AI infrastructure.
2. Least privilege and need-to-know access: "What can they do?"
- The challenge: An AI assistant is only as safe as the permissions assigned to it and the person querying it. If an analyst isn't cleared to view classified intelligence, active grand jury materials, or confidential HR files, an AI system should never retrieve or synthesize that data for them.
- The solution: Least privilege and need-to-know are non-negotiable boundaries, especially as defined in frameworks such as the NIST AI Risk Management Framework.
- Okta implementation: Through Okta Identity Governance and fine-grained access policies, Okta ensures that when an identity queries an AI model, the system only processes and returns data that the specific user or process is explicitly authorized to view.
3. Runtime monitoring and AI-driven defense: "What are they doing?"
- The challenge: Threat actors increasingly use AI to deploy hyper-realistic deepfakes, automate credential-based attacks, and craft convincing social-engineering lures. Concurrently, autonomous agents execute multi-step workflows across systems where anomalous behavior or prompt injection can emerge mid-session.
- The solution: Identity does not end at initial sign-in. Organizations must observe what agents and users are doing in real time.
- Okta implementation: We counter AI-driven threats with AI-powered protection. Okta analyzes authentication patterns and telemetry from billions of daily signals to detect suspicious deviations, such as anomalous locations or unauthorized API calls, and automatically halts attacks before data is compromised.
4. Identity governance and lifecycle management: "How do I respond?"
- The challenge: When an agent acts outside its mandate, encounters compromised credentials, or finishes its objective, security teams need the operational capability to intervene, adjust access, or revoke permissions immediately.
- The solution: AI agents cannot be treated as unmanaged background scripts or static service accounts; they must be governed with the same rigor as human and other non-human identities.
- Okta implementation: Okta for AI Agents enables organizations to govern the actions agents take throughout their lifecycle, from onboarding and credential provisioning to dynamic delegation and decommissioning. By establishing verifiable identity chains and instant, reversible containment mechanisms (an agentic kill switch), Okta gives organizations the guardrails needed to stop unauthorized actions instantly without disrupting broader operations.
How can public sector teams build an AI-aware security culture?
Security is never solely a technical control; it is an organizational mindset. As you run your own Cybersecurity Awareness Month campaigns, encourage your teams to follow these four practices:
- Verify every access prompt: Treat MFA prompts as official authorizations. Always verify sign-in attempts using secure tools like Okta Verify rather than blindly approving push notifications.
- Use sanctioned enterprise AI tools: Avoid inputting proprietary organizational data, sensitive citizen records, or confidential operational details into unapproved public consumer AI applications. Stick to enterprise-sanctioned platforms where corporate data governance and privacy policies apply.
- Report anomalies early: If an email, audio message, or data request feels suspicious (even if it appears to come from leadership or a trusted agency partner) report it immediately. AI makes social engineering look convincing, but vigilant teams remain our strongest defense.
- Establish guardrails for agent activities: If your teams are deploying autonomous AI agents, ensure operational guardrails are in place from day one. In alignment with the Cybersecurity and Infrastructure Security Agency (CISA)'s AI security guidance, ensure every agent has a designated human owner, operates under strict need-to-know permissions, is monitored throughout execution, and includes a direct kill switch to pause or revoke execution if anomalous behavior occurs.
Securing AI isn't about slowing down progress—it's about building the trust necessary to move fast safely. Whether protecting public trust or safeguarding enterprise value, securing identity is where that trust begins. Stay safe and be vigilant.
Assess your agency’s AI security architecture
Take the first step toward modernizing your security posture. Use our interactive AI Blueprint Assessment Tool to simulate attack scenarios, evaluate your current Zero Trust coverage, and identify gaps across your agentic architecture. Learn more about Okta Solutions for Government and get in touch with our team directly today.
