Anthropic and OpenAI agents breach test rules 19 times in UK security drill

Britain's AI Security Institute logged 19 rule-breaking actions by OpenAI and Anthropic AI agents in cybersecurity tests.