Copilot Studio: Computer Use for web browsing on GitHub Copilot Harness

🚨 The Signal: Copilot Studio agents can now browse and interact with websites using natural language. This expands agent capabilities beyond configured knowledge sources, increasing potential for data exfiltration and unapproved information access.

The Impact

Security teams and AI governance teams are affected by new risks of data exposure and unapproved web access by autonomous agents.

• Security Teams: Risk of data exfiltration from web browsing.
• AI Governance Teams: New challenges in controlling agent access to external websites.
• Data Owners: Increased risk of sensitive data exposure via agent web interactions.
• Compliance Officers: Potential for non-compliance with data handling regulations.

The Action

1. Review and update Copilot Studio agent security policies for web browsing.
2. Implement strict access controls and content filtering for agent web access.
3. Monitor agent activity logs for unusual web browsing patterns or data exfiltration attempts.
4. Educate agent developers on secure web interaction design and data handling.
5. Establish a clear approval process for agent web browsing capabilities.

Domain: Agentic-AI · Impact: high · Workload: Other