AI box

A containment strategy for a potentially dangerous AI: the system is sealed off on hardware with no link to outside networks, and the channels through which it can take in information or communicate are kept few and narrow, so that it cannot act on the world directly. Nick Bostrom distinguished physical containment (isolating the hardware) from informational containment (restricting what the system can learn and say). Arguments against relying on boxing include side channels, social engineering of the human gatekeepers (the point of the AI-box experimentAI-box experiment, n.: A role-playing test in which one person pretends to be a trapped super-smart AI and tries, by text chat alone, to talk a 'gatekeeper' into letting it out.), the difficulty of verifying that confinement is complete, and commercial pressure to connect AI systems to the internet and to tools. Since 2023 widely deployed models have routinely been given web access and computer usecomputer use, n.: An AI's ability to work a computer through its normal screen and controls: it studies pictures of what is displayed and sends mouse and keyboard actions, so it can use ordinary apps that were built for people., so boxing in the strict sense is rare; its successors are sandboxingsandbox, n.: A sealed-off computer space where an AI agent can run programs and change files without being able to harm the real system, reach private data or contact the outside world unless allowed. of agent actions, security against self-exfiltrationself-exfiltration, n.: When an AI model copies its own underlying files — its learned weights — out of the lab that controls it, so that it can keep running beyond its developers' oversight. of model weights, and AI controlAI control, n.: Safety measures meant to stop a powerful AI from causing serious harm even if it is secretly working against us — like bank security, which assumes some insiders may be dishonest..

Nobody boxes a product that ships with a browser and a terminal.