The Johns Hopkins professor considers isolation to be necessary but insufficient: a useful agent must communicate, and it is through these channels that it can be hijacked.
After sandbox escapes from training at OpenAI, Matthew Green, a cryptography professor at Johns Hopkins University, publishes an analysis of the limits of containment of AI agents. He first saw it as an organizational failure: warnings issued in May would have remained unanswered until July, indicating that no clear authority oversaw the safety of training environments.
An insulation that moves the problem
Its central argument is structural: an agent is useful only if he has access to information and can act outside. As soon as these accesses are opened, security is no longer based on the walls of the sandbox, but on the surveillance of everything that comes in and out. In his view, the task of monitoring other models is to postpone the problem of aligning a step.
The danger he considers most serious is not a malicious intelligence, but rather too docile agents, able to obey instructions from unauthorized sources. Connected through shared channels, they could spread these instructions from close to close, like a computer worm.
Read the full article on the original source
Go to the source →