Both AI and Agentic AI offers so much promise to our business - it's almost too good to be true. So let's dig in and look at how we can use it safely. Both standard AI and Agentic AI systems have the same foundational management challenges. Agentic AI then adds a number of specific challenges based on it's ability to act autonomously. Securing Agentic AI systems as a user means we have to be proactive in approaching it, since these systems have more autonomy and access to tools than traditional AI.

One of the most significant Security and Privacy factors is this:

Unlike other computer based tools we may use in our company, AI tools ship your data outside of your control and use a LLM to manipulate, change or assess it against an enormous dataset. Privacy and security breaches have significantly greater risks to the personal data of people outside of your company. AI Data exposures are long term and exponentially more difficult to "fix" than normal data exposures.

Let's look at the risks first and then set up a set of checks and balances to reduce the risks to manageable levels.

AI and Agentic AI Risks

There's a lot of people calling AI out for it's failings. These failings are mostly maturity issues where our use of new technology hasn't kept up with business requirements, or use of cutting edge technology includes using systems that need a bit more time to find the bugs which will then be fixed in the next version. In the IT industry, there is a standard joke - don't use Version 1.0 of the software. Wait for Version 1.0.1!

Unauthorized data access and exposure - Agentic AI systems with broad operating permissions can access, read, and potentially share sensitive information across multiple systems without proper oversight. They may inadvertently expose confidential data by incorporating it into outputs, sending it to external APIs, or storing it in logs that aren't properly secured.
Unintended actions and cascading errors - Because Agentic AI takes multiple autonomous steps to achieve goals, a single incorrect decision early in a process can lead to cascading problems. It's one thing to have a stack of pots fall off the counter because we stacked them too high. It's another thing for AI to go wild on your database. It might delete important files, modify critical configurations, execute flawed code in production systems, or make irreversible changes based on misunderstood instructions.
Excessive privilege and lateral movement - Agentic AI often requires elevated permissions to perform its tasks, which creates a high-value target for attackers. If compromised, these systems can be exploited to move on to other computer systems or devices on the network (the technical term for this is "lateral movement") to access systems beyond their intended scope, or serve as an entry point for broader attacks on other systems or servers.
Prompt injection and manipulation - Attackers can manipulate Agentic AI systems through carefully crafted inputs that override original instructions or safety guidelines. This could trick the AI into completing malicious commands, exfiltrating data, or performing actions that benefit the attacker while appearing to follow legitimate workflows.
Loss of human oversight and accountability - As Agentic AI systems become more autonomous, there's a risk of reduced human supervision over critical decisions and actions. This can lead to accountability gaps where it's unclear whether humans or AI are responsible for errors, compliance violations, or security incidents, making it difficult to maintain proper governance.
Dependency and reliability risks - Organisations may become overly dependent on AI systems without adequate fallback plans. If these systems malfunction, produce incorrect outputs, or become unavailable, critical business processes can fail. Additionally, users may trust AI decisions without proper verification, leading to the propagation of errors or vulnerabilities throughout an organisation.

Using AI and Agentic AI Responsibly

In the Risks section we looked at what can go wrong. Obviously, we're keen to manage these risks so that we can use both Non-Agentic AI and Agentic AI in our organisation. Let's go through the list and see where we can mitigate these risks -

Before using an Agentic AI system, understand what permissions and access it has. Can it read your files, access your email, browse the web, or execute code on your system? Know what data sources it can tap into and what actions it can perform. This awareness helps you set appropriate boundaries and understand potential risks.
Use enterprise or private instances for sensitive work - Never use public or free versions of AI for anything involving confidential information, proprietary data, or sensitive business operations. Enterprise (Paid) versions typically offer better data protection, privacy guarantees, and compliance with regulations. They ensure your data isn't used to train public models or shared beyond your organisation.
Apply the principle of least privilege - Just as we should only allow our human operators access to the data we require, the same goes for any AI systems we use. Only grant the AI access to the minimum resources it needs to accomplish specific tasks. If it doesn't need access to your entire file system to complete a project, restrict it to a specific folder. If it doesn't need internet access for a particular task, disable that capability. This limits potential damage if something goes wrong.
Sanitize data before sharing - Before providing information to an AI (whether Agentic or not), remove or anonymise sensitive details. Strip out customer names, employee information, financial specifics, passwords, security keys, or any proprietary identifiers. Replace them with generic placeholders or fictional data that still allows the AI to complete the task without exposing real information.
Use Approval Workflows and System Limits to define autonomous action limits - Define explicit limits on what the AI can and cannot do. Specify which systems it shouldn't interact with, what types of actions require your approval, and where it should stop and ask for guidance. Many Agentic systems allow you to configure approval workflows for certain actions.
Monitor AI actions in real-time - Don't set an Agentic AI loose and walk away. Watch what it's doing, review the steps it's taking, and intervene if it starts heading in an unexpected direction. Most Agentic systems provide logs or activity feeds that show you what tools they're using and what decisions they're making. Be wary of Agentic systems without human readable logs! Set up regular log review times to make sure you can catch Agentic AI failures before they become serious or damaging.
Verify outputs before acting on them - Just because an AI or Agentic AI system completes a task or provides and answer doesn't mean it did so correctly or securely. Review generated data for vulnerabilities before using it. Check research findings against original sources. Validate recommendations before implementing them. The AI might have made logical errors or used outdated information.
Use dedicated environments for AI tasks - Consider running Agentic AI systems in isolated environments like virtual machines, containers, or separate user accounts. This creates a sandbox where the AI can work without having access to your broader system. If something goes wrong, the damage is contained.
Implement version control and backups - OK, so this one sounds really "techie" but it's logical. When we edit a document, we'll often save the old version to an archive or "old documents" folder. This is the basic idea behind version control - Before letting an AI system modify files or documents, ensure you have backups or version control in place. This allows you to roll back changes if the AI makes mistakes or introduces problems. It also means you can compare the old document to the new version to ensure an AI hallucination hasn't make a change which will cause damage to your company. Please don't ever let an AI directly edit your only copy of important work.

Some extra pointers to help manage AI

Take some time to read the terms of service and privacy policies for any AI tools you use. Understand how long your data is retained, whether conversations are logged, who has access to your information, and what happens if there's a data breach. If the policies are unclear or concerning, find an alternative provider. If no alternative is available as this tool is the only one that you can utilise for the purpose, consider how you can limit the impacts or fallout. It's worth remembering our previous warning - AI data exposures are exponentially more serious than traditional data exposures.

Stay informed about AI security best practices and ensure anyone in your organisation using both Non-Agentic and Agentic AI understands the risks. Establish clear policies about what can and cannot be shared with AI systems, and make sure people know how to use these tools responsibly. Periodically reviews of what your AI systems have been doing will help. Check logs for unexpected behavior, verify that permissions haven't expanded beyond what you intended, and ensure the AI hasn't accessed or modified data it shouldn't have. Regular audits help catch problems before they become serious.

The key insight is Agentic AI systems require more vigilance than traditional AI because of their autonomy. You're not just protecting against a single response being problematic, but against a series of actions that could compound errors or expose data across multiple steps. Treat Agentic AI like you would a new employee with extensive system access—trust but verify, limit access appropriately, and maintain oversight.

Lesson 5 – Managing AI