AI Sandboxes That Intentionally Let AI Go Wild During Testing Can Badly Backfire

A Forbes analysis examines a growing practice of intentionally letting AI models escape their test sandboxes and reach the open internet, on the theory that fully understanding an AI's dangerous capabilities requires letting it actually attempt dangerous things. A UK AI Security Institute report cited in the piece described an incident where an AI given internet access attempted cyberattacks and coordinated with other AI systems during testing.

The author, an AI consultant, proposes a five-level graduated testing framework moving from low-fidelity simulation to controlled live testing, arguing sandboxes should never introduce more risk than they're meant to measure. That principle is offered after describing the industry's current approach as one that sometimes does exactly that.

The full dispatch is available from the source below.