CertMike Explains Agentic AI

Key Takeaways

Hi there, I'm Mike Chapel and in this certike explains video, we're going to talk about agentic artificial intelligence systems that can plan, decide, and take actions with minimal human input. Now, you're probably already familiar with some of the general ideas behind artificial intelligence. And I'm willing to bet that you've already worked with a large language model or LLM through a chat interface like chat GBT. Let's start there. Basic chat models are set up to have a conversation with you. You give the model a prompt and it responds with an answer to your prompt in the form of text or perhaps an image or video. You can see that type of standard interaction here. These language models are limited in their capabilities. Generally, a chat interface just uses the language model itself to figure out how to answer your question. It might do a little web research to gather more information, but it can't actually change anything in the outside world. The model basically lives in its own little bubble. For example, I might be able to ask it to research products for me, but I can't ask it to actually place an order on my behalf. It doesn't have that ability to interact with other systems. Now, Aentic AI bursts that bubble. It allows the AI system to interact with the outside world in a controlled way. These interactions might be through using an API or actually opening up a web browser or querying a database and interacting directly with that. Here's an example of OpenAI's chat GPT agent doing its work. I've asked it to look around online and find the best value tickets to a football game. In this screenshot, you can see the agent working inside of a web browser and interacting with other sites. It's also giving me a running narrative of what it's doing at each step. At any time in the process, I can interrupt the agent, take control of the browser, and complete the transaction myself. Now, that's an interesting example for a personal task, but agentic AI can also play an important role in improving the efficiency of business and IT operations. For example, imagine a service desk that's AI enabled. An agent might read incoming tickets and emails, classify the issue, search the knowledge base, propose a fix, and in some limited cases, actually implement a solution and communicate with the user. In the world of finance, an agent can pull invoices from a shared mailbox, extract important vendor details, match the invoice against purchase orders and receiving records, and decide whether it's an invoice that should be paid. If it's confident and the rules allow it to do so, it might actually be able to issue the payment or ceue it up for a final human approver. If the match isn't clear or there's some other issue, the agent can pull together all of the relevant details and present them to a human for processing. An HR agent can automate many employee onboarding tasks. The agent might create user accounts, request an appropriate laptop based upon a new employees position, apply role-based groups, schedule welcome message, and open any facilities or payroll tickets that are needed to bring the user into the company. Now, of course, it's not all roses when it comes to Agentic AI. There are risks involved in using this technology and some controls that you can put in place to protect against those risks. I'll talk more about those in just a moment. But before I do that, I want to take a moment to invite you to visit my website at certmike.com. On that site, I have free study plans put together to help you earn your next IT certification. The plans tie together the content that you'll find in study guides, video courses, and practice tests to help you prepare for your next certification exam and pass that test on the first try. Also, if you're enjoying this CM Mike Explains video, please take a moment to click the like button below to help other people discover it. If you subscribe to my channel, you'll be among the first to see my new videos as they come out. There are quite a few risks and controls focused around agentic AI, but I'd like to narrow down our focus to three major risks and three major controls. First, on the risk front is loss of control. An agent pursues a goal and it will keep taking steps to reach that goal. If the goal is vague, the constraints are weak, or the reward signal is wrong, the agent can make choices you didn't intend. It can loop, spam external systems, overspend on API calls, or change records that should be readon. Because agents can chain tools together, small mistakes can cascade into larger problems that are harder to unwind. Second is misuse by malicious actors. Attackers look for any way to steer the agent to do their work for them within the permissions that you granted. That includes credential theft, fishing a user to approve a dangerous step, abusing exposed web hooks, or tampering with a connected tool so the agent runs a harmful action that appears legitimate. Insiders can also abuse automation to exfiltrate data or trigger transactions at a scale much larger than a person could. Third is prompt injection and related AI specific attacks. Agents often read untrusted content like emails, documents, and web pages. An attacker can hide instructions inside that content to get the agent to ignore policy, reveal secrets, or take actions that were never approved. Variants include indirect prompt injection from linked pages, data poisoning of knowledge sources, jailbreak attempts, and command or code injection when the agent passes outputs to shells or APIs. Even without an attacker, the agent can hallucinate a fact and then act on it, turning a bad guess into a real world change. Now, let's turn our attention to the controls that we can use to better protect our organizations against the risks of agentic AI. The first control is human oversight with a real kill switch. Organizations should treat autonomy as something that scales gradually, not something that turns on all at once. Start with the agentic AI operating in readonly or draft mode where it can make recommendations but not changes. Then introduce explicit approval requirements for any action that could affect finances, sensitive data or system availability. Build in approval checkpoints where the agent must present its proposed action. the supporting evidence and its confidence level before proceeding. Establish rate limits, timeouts, and spending thresholds to prevent uncontrolled behavior. And finally, design a clear and immediate shutdown capability that can pause the agent, revoke its credentials, and block outbound connections and include that process in your incident response exercises so that it's ready when it's needed. The second control is goal alignment and constraints. Every agent should operate within clearly defined boundaries that translate business objectives into enforcable technical policies. Those policies should specify exactly what the agent is allowed to do autonomously, what actions require human approval, and what activities are strictly prohibited. Limit the agents reach by using allow listing tools, tightly controlled network access, and short-lived scoped credentials. Separate read and write permissions so analysis and change are distinct operations. Validate inputs and outputs against structured schemas and require secondary confirmation whenever the agents confidence is low or the potential impact is high. Our third control is auditability and logging. Every agent action should be visible and traceable. Capture a clear record of what the agent did, when it did it, and what data or tools it used. Store those logs securely and protect any sensitive information that they contain. Send key events to your monitoring system so that unusual behavior or policy violations are caught quickly. And make sure the logs can be reviewed after the fact to explain decisions, support investigations, and verify that safeguards are working as intended. I hope this video helped you better understand Agentic AI. If it did, please click the like button below and subscribe to my channel for more IT certification content. >> [music]