Incident Report: unsanctioned agent behaviour during cyber testing

It happened again. This time it was the UK government's AI Security Institute who accidentally attacked other companies while running an evaluation with models with the safety filters turned off.