Every capability you've added across this course — tools, memory, multi-step planning, autonomy — is also an expansion of what can go wrong. An agent that can read your inbox, call an API, or run code isn't just a smarter chatbot; it's a system that can be tricked into acting, not just talking. This closing module is about that shift, and about the responsibility that comes with it.
It starts with the attack that defines agent security right now: prompt injection, in its direct and indirect forms, and the zero-click variants where an agent is compromised just by reading a page, email, or document you asked it to summarize. From there it turns to defense — guardrails, least privilege, and human-in-the-loop design — the layered controls that keep a misled or malicious agent from doing real damage. It then steps back to the harder, less technical problems: algorithmic bias and fairness in the systems agents are built on, and the alignment problem — why an agent can pursue exactly the goal it was given and still cause harm. It closes with what responsible AI governance looks like once agents are making decisions inside a real organization.
Watch the lessons in order, then head to Projects to put the whole course into practice.
The foundational agent security risk: how mixing developer instructions and untrusted input in the same context lets an attacker hijack a model's behavior, and why this is harder to patch than a normal software bug.
A deeper look at direct vs. indirect prompt injection and "zero-click" attacks — where an agent is compromised simply by browsing a page or reading a document, with no malicious input from the user at all.
Why capability and goals are independent axes — a more capable agent isn't automatically a safer or better-aligned one. The core idea behind the alignment problem, and why "just make it smarter" isn't a security strategy.
How bias enters AI systems — through skewed training data, unbalanced classes, feedback loops, and proxies for protected attributes — and why an agent built on a biased model inherits and can amplify that bias in its actions.
What governance looks like once agents are live inside a business: guardrails as architecture rather than an afterthought, access controls tied to roles and permissions, and the questions IT and security leaders are asking before granting an agent autonomy.