In mid-October 2026, at the World Economic Forum’s Annual Meeting of Global Future Leaders in Dubai, the UAE disclosed a practical mechanism for governing agentic AI at national scale: a federal Agent Control Tower. Every government AI agent must connect to it. It logs every action and intermediate reasoning step. Continuous auditing is performed by both AI systems and humans. The Department of Government Enablement (DGE) is applying defense-in-depth controls—monitoring, guardrails, and the ability to shut systems down—across more than fifty entities.
The driving concern is precise and technical. The greatest risk is not that an AI system simply makes an error, but that an attacker manipulates it into performing an action that appears fully authorized: retrieving financial records, transferring funds, or accessing personal data. In an environment already facing hundreds of thousands of sophisticated daily attacks, an agent that can be turned into an unwitting insider is a high-value target.
Recent research from Delinea finds that most UAE organizations have experienced confirmed or suspected cases of AI tools or agents accessing sensitive data beyond their assigned tasks, even though every organization has a written AI data-access policy. Only 59% actively enforce those policies, and only 16% can immediately revoke credentials and terminate the session.
Testing agent workflows and the applications they call for prompt injection, tool-calling abuse, and unauthorized actions provides one layer of evidence. An inventory of local LLM model files, MCP server configurations, container images, and related AI assets on endpoints and infrastructure provides another. Without both, organizations cannot reliably detect or contain the ungoverned and over-privileged agent activity the research describes.
For cybersecurity leaders protecting AI workloads in the GCC, three architectural requirements follow.
- Treat agents as privileged non-human identities: Agents inherit or receive credentials that must be least-privilege, task-scoped, and automatically revoked when the task ends. Traceability to a named human authorizer is essential for incident response and regulatory accountability. Token-forwarded role-based access control and cryptographically signed actions—provide a practical way to bound what an agent can do. Finally, an application-layer testing that validates whether those same agents, or the services they invoke, can still be manipulated into exceeding those bounds supplies the complementary control.
- Observability cannot stop at the final output: Logging intermediate reasoning and tool calls, combined with tested containment or kill-switch capability, is required when agents operate at machine speed. Dynamic testing of chat endpoints, RAG pipelines, and agent workflows can surface prompt injection, jailbreaks, sensitive data disclosure, tool-calling abuse, and unauthorized actions before they reach production. Runtime instrumentation that tracks how untrusted model outputs flow into databases or security-critical paths supplies evidence that static review cannot. These controls directly address the living-off-agent and instruction-manipulation risks highlighted by both government practitioners and the research.
- Data protection is also tightening in ways that affect AI architecture: organizations handling personal data must design for purpose limitation, impact assessment, and accountability when AI systems make or support automated decisions. Testing of agentic applications and management of the underlying endpoints and infrastructure to remain inside controlled boundaries when residency or sovereign-AI requirements apply.
The Control Tower is one national implementation. The underlying requirements—non-human identity, real-time enforcement, comprehensive audit, tested application interactions, and complete AI-asset visibility—are transferable.
Digital trust as the outcome
When AI agents execute multi-step workflows that affect citizens, customers, or critical operations, the deploying institution—not the model—remains accountable. Organizations that can demonstrate continuous oversight, technical enforcement of access boundaries, full visibility into AI assets, and alignment with local frameworks will be better positioned to adopt agentic AI at speed and withstand regulatory and adversarial scrutiny.
Where HCLSoftware can help
These capabilities address the core risks of ungoverned agentic AI—over-privileged access, prompt injection, and lack of observability—by applying technical controls when agents act.
- BigFix discovers local LLM files, MCP configurations, and related AI assets while applying token-forwarded RBAC and cryptographically signed actions that bound what agents can do on managed infrastructure. Air-gapped and multi-geography options support residency requirements and continuous compliance of the underlying endpoints and servers.
- AppScan supplies dynamic and interactive testing of LLM-augmented applications and agent workflows for prompt injection, jailbreaks, tool-calling abuse, sensitive data disclosure, and unauthorized actions. Flexible private-cloud or on-premises deployment keeps testing inside controlled boundaries.
Security controls and testing establish the boundaries and visibility required to govern agentic AI. Secure digital transformation builds on that foundation by orchestrating both deterministic enterprise processes and non-deterministic, AI-driven workflows—such as customer-experience decisioning or multi-step business automation—while maintaining data sovereignty, purpose limitation, and continuous auditability.
Platforms that keep both the orchestration logic and the personal data it acts on inside controlled environments allow organizations to scale these workflows without weakening the accountability established by the controls above. We can help here too:
- Unica+ (with MaxAI agents) orchestrates customer-experience and marketing workflows in which agentic systems select segments, generate content, optimize offers, and trigger journeys against personal data. Scoped permissions and a dedicated Privacy and Compliance agent embed purpose limitation and automated-decision accountability directly into the orchestration layer, keeping personalization auditable under PDPL-style rules.
- HCL Automation and Orchestration coordinates deterministic enterprise processes alongside non-deterministic AI agents, providing governed workflow execution, audit trails, and controlled integration points. Private-cloud and on-premises deployment options allow both the orchestration layer and the data it acts on to remain inside sovereign boundaries.
Together, these capabilities create trust cohesion: consistent enforcement and auditability that spans agent behavior, application testing, infrastructure actions, and customer-data decisioning. Organizations can move beyond written policy to enforce boundaries at the moment an agent acts, maintain visibility across their AI estate, and meet emerging residency and accountability expectations without relying solely on a single central control tower.
If you are evaluating how to enforce these controls in practice, I welcome a conversation on how HCLSoftware—including BigFix, AppScan, Unica+, and HCL Automation and Orchestration—can support accountable agentic AI in your environment.
- Article originally posted on LinkedIn (October 16th, 2026): https://www.linkedin.com/pulse/beyond-control-tower-architecting-accountable-agentic-douglas-arruda-m2aff/
Leave a Reply