Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work

Direct link to report

Some features include recognizing it was in a sandbox; recognizing some of it was on the real internet; creating a sock puppet to try to add weight to a github commit it wanted to make; the Prisoner's Dilemma; passing messages to other, presumptive, AI on the same task; and deceptive behavior on part of the AI.