Agentic AI and New Security Challenges

Artificial intelligence in business is evolving. While many AI systems today focus on assisting users, agentic AI introduces a new level of capability.

AI systems can now perform the following:

1. Interpret natural language instructions
2. Reason about context
3. Plan next steps
4. Execute actions across systems

This means AI is no longer limited to supporting tasks; it can orchestrate actions across business processes.

Traditional enterprise security was designed for a world where humans make decisions and systems execute them. Security focused on questions such as:

• Who has access to what?
• Which systems are exposed?
• Where is data stored or transmitted?

These questions remain important, but with agentic AI, they are no longer sufficient. Organizations must now also consider:

• What the agent is being told
• How it interprets instructions
• Which tools it is allowed to invoke
• Whether it should act autonomously

In this context, poorly governed prompts or instructions can introduce new security risks. This represents a fundamental shift: security is no longer only about protecting systems; it is also about governing how AI-driven decisions and actions are executed.

As AI systems gain the ability to act across systems, the potential impact of errors or misuse increases significantly. A traditional system issue may affect a single application or process. An AI-driven action, if not properly controlled, can extend across multiple systems and workflows.

To address this, additional controls are required, including:

Content filtering

To detect unsafe or manipulated inputs.

Data masking

To reduce exposure of sensitive information.

Structured prompt handling

To guide interactions effectively.

Strict limits

Defining which tools and actions can be invoked.

Human approval

Required for high-risk or irreversible actions.

Auditability

To ensure traceability and accountability.

These controls are no longer optional; they are becoming core design requirements for enterprise AI.

SAP applies a layered security model, where multiple safeguards work together across the full lifecycle of AI interactions. This includes:

• Governance and compliance controls
• Data and context protection
• Model and inference safeguards
• Runtime execution controls
• Orchestration and policy enforcement

No single control is sufficient on its own. Only the combination of these layers ensures that risks are effectively managed.

The goal of securing agentic AI is not to eliminate autonomy, but to constrain it within defined boundaries. SAP ensures that:

• AI operates within user authorization contexts
• Actions are limited through scoped permissions
• Tool access is controlled and governed
• High-impact decisions can require human approval

This ensures that AI systems remain aligned with business rules and organizational control, even as they become more capable.

Agentic AI enables new levels of efficiency and automation. But as capability increases, so does the need for control. By combining layered security, governance mechanisms, and continuous oversight, SAP enables organizations to scale AI adoption while maintaining security, compliance, and accountability.

Agentic AI introduces a new level of capability by enabling AI systems not only to assist users, but also to reason, plan, and execute actions across business processes. Because of this, security must go beyond protecting systems and also govern how AI-driven decisions and actions are performed.

SAP supports this through:

• Layered controls such as filtering, masking, orchestration, and policy enforcement
• Restrictions on tool access and execution boundaries
• Human approval for high-risk or irreversible actions
• Continuous monitoring, logging, and traceability

These safeguards help ensure that agentic AI remains aligned with business rules, security requirements, and organizational control.