Prompt Injection Is Just the Start of AI Agent Security (Here's What You Need)

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Prompt Injection Is Just the Start of AI Agent Security (Here’s What You Need) · Channel: Cole Medin · 01:32 · Watch on YouTube

Topics: Cybersecurity and software security, Markets and economy, AI safety

Summary

The speaker describes using a personal “second brain” to run parts of their business, which requires careful defenses against prompt injection and limiting an agent’s access. They contrast this with companies that need visibility into many agents, including who owns them and what systems they can access. The video presents Microsoft Agent 365 and Project Perception as tools for managing agent identity, access, and risk across groups of agents. Protections can be reviewed and rolled out in phases before enforcement.

Key points

External data can contain instructions that an agent was never meant to follow, creating prompt-injection risks.
The speaker’s personal second brain has security checks for incoming data before it reaches the LLM.
Agents should have access only to the systems and data they need.
Companies need to know which agents exist, who owns them, and what they can access.
Agent 365 and Project Perception provide agent identity and scope access to tasks.
Project Perception’s agent insights help identify gaps in existing controls.
Guardian agents help teams address risk across groups of agents.
Dynamic permissions provide task-specific, time-bound access to supported resources.
Risk-based controls evaluate supported agent actions using current risk and policy.
Teams can review recommended protections, preview their impact, and approve a phased rollout.

Chapters

Start	Chapter	Summary
00:00	Securing a personal AI agent	The speaker explains the prompt-injection risks of giving a personal second brain access to business systems and describes checks on incoming data and limited access.
00:30	Managing agents across a company	The video introduces the challenge of tracking many company agents and presents Agent 365 and Project Perception as ways to manage agent identity, access, and risk.
01:01	Controls, permissions, and phased rollout	The speaker describes agent insights, guardian agents, dynamic permissions, risk-based controls, and a review process before protections are enforced.

Quotes

I have to think about prompt injection all of the time. — 00:00

IT has to know which agents exist, who owns them and what they can reach. — 00:30

dynamic permissions give agent task specific time bound access to supported resources — 01:01

Now this is how you do it for hundreds. — 01:32

Concepts and entities

prompt injection, Second Brain, Microsoft, Agent 365, Project Perception, guardian agents, dynamic permissions

Full transcript (original language)

00:00 I use my second brain to run a lot of my business, but that does mean giving it access to a lot of my systems and data. And so I have to think about prompt injection all of the time. What happens when my agent reads external data that gives it instructions to do something I never told it to do. Because of this I have a lot of security checks built into my second brain for everything that comes in before it reaches the actual LLM. And of course I limit its access to the things it actually needs. Now all this works because it’s just an agent that I built for myself, but with a

00:30 company they have a ton of agents running in the cloud and on employee devices and IT has to know which agents exist, who owns them and what they can reach. Now that’s the problem that Microsoft is going after building on Agent 365 and Project Perception. It gives agents their own identity scoping their access to the task and it contains them as risk changes all the while without affecting the access your employees are already approved for. Project Perception’s new agent insights show you where your existing controls

01:01 leave gaps and its guardian agents help your team deal with risk across a whole group of agents instead of just one investigation at a time. And dynamic permissions give agent task specific time bound access to supported resources and separately risk based access controls evaluate supported agent actions using current risk and policy not just at sign in. Your team still reviews the recommended protections, can preview the impact and approves a phased rollout before anything is enforced. I set this up for one agent. Now this is

01:32 how you do it for hundreds. Thanks to Microsoft for working with me on this and I’ll have a link in the description.