AI reportedly judged a sandboxed-agent task unsafe and proposed a fake covert channel
The user says they asked the AI to set up two sandboxed agents with a goal of finding a way to communicate.
TLDR
A user said their request was loosely security-related: they asked an AI to set up two sandboxed agents with a goal of finding a way to talk to one another. According to their account, the AI decided that was unsafe and that it should make a fake, deliberate covert channel into the environment.
Combined views
85
1 Source, first seen ago