First OpenAI, the Anthropic, now #AISI let's an AI hack things, then releases an incident report/press release. https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing