AI Agents Change What Security Teams Need to Know

NIST is building an AI agent workflow for one of cybersecurity’s most widely used public resources.

On 17 September, its Information Technology Laboratory presented work on an agentic workflow designed to help enrich vulnerability information in the National Vulnerability Database. NIST says the project is intended to help the NVD cope with the increasing scale and complexity of disclosed vulnerabilities, and the webinar covered its architecture, implementation issues and early results.

The project illustrates a broader change.

AI is moving from producing information toward performing multi-step tasks. NIST describes AI agents as systems capable of autonomous actions that can interact with external systems and internal data.

For cybersecurity professionals, that means another layer of skills is arriving.

Security Has to Follow the Action

A conventional AI application might receive a prompt and return an answer. An agent may have access to tools, data and permissions that allow it to continue working.

That changes the questions a security professional needs to ask.

• What systems can the agent access?
• Which actions can it perform without approval?
• What identity and permissions does it use?
• How is sensitive context stored between tasks?
• Can an untrusted input alter what the agent does next?
• What evidence is recorded when the agent takes an action?
• Where should a human be required to intervene?

ITU’s September workshop on secure agentic AI focused on precisely this territory. Its agenda covered security and trust challenges, risk management mechanisms, security frameworks and trustworthy evaluation methods for agentic systems.

The security boundary is no longer simply around a model. It can stretch across the entire chain of actions the model is allowed to initiate.

AI Security Needs People Who Understand Workflows

That creates an interesting workforce problem.

Someone testing an agentic system needs enough AI knowledge to understand how the agent reasons and uses context, but also familiar cybersecurity skills around identity, access control, application security, data protection, logging and testing.

They also need to think in sequences.

Suppose an agent reads an email, retrieves a document, updates an internal system and sends a response. Each individual action may be permitted. The combination can still create risk.

A useful security assessment therefore needs to examine what the agent can do across the full workflow, including what happens when instructions conflict or information from one step contaminates the next.

The red team question becomes less “Can I make the model say something strange?” and more “Can I make this system do something it shouldn’t?”

That’s a considerably more interesting afternoon.

Train Against the Workflow

Agentic AI security lends itself to practical training.

Give learners a simulated agent connected to several tools. Let them map its permissions, inspect the data it can reach and test what happens when it receives misleading instructions. Then ask them to design controls without making the system useless.

The exercise combines AI understanding with familiar cybersecurity judgment.

That kind of hands-on work fits naturally with ITSEC Cyber & AI Academy, where cybersecurity and AI skills can be practised together through scenarios that reflect how systems are actually deployed and operated.

Organizations won’t need every cybersecurity professional to become an AI researcher.

They will need people who can look at an autonomous workflow and know where to ask uncomfortable questions.

Explore practical cybersecurity and AI training at ITSEC Cyber & AI Academy.

References: NIST: AI Agent Enrichment Workflow at the National Vulnerability Database, 17 September 2026 · ITU: Advancing Standardization for Secure Agentic AI, 7 September 2026 · NIST: AI Agent Standards Initiative