An agent that can search the web, call tools, and take actions can also do the wrong thing, whether by mistake, by retrying forever, or because someone tricked it. Safety is the work of making sure an agent stays inside the boundaries you set, and guardrails are the checks and limits that enforce those boundaries while it runs.
The lessons start with a quick overview of guardrail types: limits on steps, time, and cost, plus input, action, and output checks, and when to require human approval. Then you'll look at prompt injection, the best-known attack on LLM applications, where crafted text overrides an agent's instructions, and why it is so hard to fix completely.
As you watch, notice a recurring idea: don't rely on instructions alone. Build the system so unsafe actions are blocked by design, using least-privilege permissions, checks before tool calls, and approval for high-risk steps like payments or deletions. Layered defenses matter because no single guardrail is perfect.
A quick visual overview of agent guardrails: limits on steps, time, tokens, and cost; input, action, and output checks; least-privilege access; and when an agent should ask for human approval.
Video 3minA clear walkthrough of why prompt injection works: how LLM apps mix instructions and untrusted input, how this differs from classic injection attacks, and why preventing it is difficult.
Video 13minAn introduction to the security weaknesses specific to AI and LLMs, why attackers can manipulate them with plain words instead of code, and why securing AI calls for a new mindset.
Video 15min