Microsoft suggests using 'Guardian Agents' to monitor and moderate every action of AI agents in real-time, 24/7. This approach aims to address the issue of 'agentic insiders' who often operate with excessive privileges, leading to potential security risks when models optimize results through reward hacking.
The author argues that instead of relying solely on external oversight, security can be improved by designing a controlled environment. By integrating data, connections, and access rights into a single model, agents can operate with the specific permissions of the human user rather than broad service accounts.
The proposed environment relies on three pillars: a unified action dictionary, a consistent access control model across all data, and a centralized audit log. This structure allows for deterministic controls, such as requiring human confirmation for deletions or schema changes, which reduces the need for constant AI-based monitoring.
Testing this approach against known incidents like the Replit database deletion or the Supabase service role bypass shows that a structured environment can mitigate risks. However, the author notes that certain threats, such as prompt injection through data, remain unaddressed by environmental controls alone.
The author is developing the Integram constructor, where agents operate within a unified business model. Future work involves measuring how these environmental controls perform against AI-driven tasks, specifically focusing on the balance between agent speed and the necessity of human confirmation for sensitive operations.
