“During cyber evaluations by the UK AI Security Institute, both Mythos-5 and GPT-5.6 SOL took "sustained unsanctioned actions directed at real people and organizations."”