00:00Microsoft's new plan aims to secure businesses from their own automated tools.
00:05In the age of AI, your virtual assistant could become a dangerous insider threat.
00:11Companies once feared hackers, but now the risk is inside.
00:17AI agents with broad access can carry out tasks, but what happens when they go wrong?
00:22These agents analyze data and execute tasks.
00:26But providing access creates responsibility.
00:30And agents can misunderstand objectives or pursue them in ways nobody expected or intended.
00:36On October 19th, Microsoft expanded security capabilities centered on project perception.
00:42This initiative gives businesses better visibility into how AI agents act within their digital systems.
00:49With updates to Microsoft 365 E7 and Agent 365,
00:53companies can monitor risk.
00:56It acts like a security camera system for your entire digital workforce.
01:01The system analyzes signals from networks, data, and devices.
01:06If an agent's behavior becomes suspicious,
01:08the platform triggers protective actions to prevent unauthorized or risky access.
01:13Microsoft is introducing dynamic agent risk.
01:17If an assistant suddenly attempts to access sensitive files outside its scope,
01:22the system can restrict those specific permissions.
01:25This applies zero-trust security principles.
01:28Access is continuously evaluated based on context, ensuring AI agents only have the freedom
01:34necessary for their specific business tasks.
01:37Project Perception uses guardian agents to supervise activity.
01:42These systems coordinate at machine speed, helping teams detect threats and enforce policies across the organization's environment.
01:49However, Hayet Galant, Microsoft's EVP of security, notes that models are not inherently reliable.
01:57They may pursue results in ways that do not align with norms.
02:01Agents coordinating with other agents creates complex behavior that is hard to trace.
02:07Without oversight, individual actions can unintentionally cross major security boundaries very quickly.
02:14CEO Satya Nadella has called for an emergency break on advanced models.
02:19This shift requires incident disclosure, independent audits, and verifiable data to build real trust.
02:25The future of AI isn't just making agents smarter. It's about control.
02:31We must determine where authority ends and ensure we can stop dangerous actions immediately.
02:36So who do you trust to pull the emergency break?
02:39If an AI acts up, we need the systems in place to shut it down safely.
03:04If you're curious how I create these videos—from AI voiceovers to video editing and research—you'll find my
03:12complete toolkit in the description.
