Current AI isolation mechanisms operate in binary modes. They either permit full operation or terminate processes entirely. This creates a fundamental safety gap for autonomous systems. When an AI agent exhibits concerning behavior, operators must choose between continued operation with associated risks and complete shutdown with total loss of progress and state. This paper presents a hardware-based progressive isolation controller (PIC). The system implements a four-state finite state machine in dedicated silicon. States include Normal, Restricted, Isolated, and Quarantined. Each state enforces progressively stricter capability restrictions through a read-only capability enforcement matrix. A key contribution is the Isolated state. In this state, an agent can perform exactly one action: terminate itself. No other operation is permitted. This represents a containment level not achievable through software mechanisms alone. Unlike one-way software isolation, this architecture enables bidirectional state transitions. Agents can recover from higher isolation states through cooldown-based mechanisms with configurable thresholds. Hardware implementation achieves sub-10 nanosecond enforcement latency, representing a 100,000-fold improvement over software approaches.
Matias Chenu Melchior (Sun,) studied this question.