Secure Personal Agent is research into a runtime control layer for tool-using agents: task-specific authority, short-lived credentials, and policy checks before execution.
Research & development · Not commercially available
This work aims at autonomy inside defined boundaries: task-specific, least-privilege authority, with each proposed tool call evaluated against identity, policy, and context before execution.
Task-specific, least-privilege authority instead of standing access — read:tickets, not root.
The proposed approach evaluates each tool call against identity, policy, and context before execution.
The design intent is that actions exceeding the approved boundary are blocked, constrained, or escalated, not only logged.
Secure Personal Agent is part of mace AI’s Deep Model Trust research and development. It is not commercially available.
Explore the research → Assess your current agent with Fisher