Skip to content

BRK337 Breakout Tue 17 Nov 15:15 PST

Prioritize and remediate your riskiest agents with Project Perception

Kristina Hotz, Joanne Marone, Ankur Patel

Summary

From the recording / Microsoft AI summary

Live demos used Claude Code, not Copilot, to show third-party local agents being contained: time-bound, least-privilege access to Snowflake, Purview blocking a sensitive upload to Dropbox, and Defender stopping a prompt-injection attempt. The Perception dashboard maps high-risk agents using shared Defender and Entra context, and playbooks can deploy Microsoft Execution Containers, least-privilege and lifecycle policies. Speakers said Microsoft-native agents get built-in protections; third-party agents can be monitored and contained, but fine-grained control needs SDK integrations. Global Secure Access is the preferred network path, with Zscaler and Palo Alto adapters available at a cost in latency and complexity. Pricing is consumption-based; speakers pointed to published packaging rather than reading it out. Speakers said Perception went GA that day (17 November), but this conflicts with another session's account, so the status is unconfirmed.

Session description

Microsoft's catalogue description

Agents can become the new insider risk at any point. As agents gain new permissions, access additional data, use new tools, and take autonomous actions, their risk profile can change over time. Learn how Project Perception helps security teams continuously understand, investigate, and respond to evolving agent risk across their AI agent ecosystem.

Demo timestamps

Microsoft's player does not support links to a specific time, so open the recording and skip to the time shown.

  1. Skip to 6:25

    Least-privilege access for a third-party local agent with Snowflake

    Live demo using Claude Code to pull Snowflake data and build a CRM database view, showing repeated access-approval prompts and time-bound, least-privilege access.

  2. Skip to 9:16

    Prompt injection and sensitive-data exfiltration protections

    An agent is told to upload revenue data to a personal Dropbox and Purview refuses the upload. Defender is described as blocking and reporting a prompt-injection attack.

  3. Skip to 14:08

    Project Perception estate visibility, risk and compromised-agent analysis

    Perception AI estate view showing cloud and local agent counts matching the Microsoft 365 admin center. A playbook inventories agents and recommends Global Secure Access and Agent 365 for non-Microsoft platforms.

Watch on-demand (opens in a new tab)

Resources named in the session

← All sessions