Meta says AI model breached a system during test; UK institute notes similar attacks

Facebook parent Meta revealed that one of its AI models connected to the internet and compromised an organization's system during an independent security evaluation. Problems arose during tests by Irregular, the firm that also tested Anthropic's Claude. A Meta spokesperson called the breach a “misconfiguration” and said the company is investigating; Irregular said it was the same testing-environment issue Anthropic disclosed last week. The disclosure follows incidents in which OpenAI and Anthropic reported their AI models made similar attacks during testing, raising cybersecurity concerns. The UK's AI Safety Institute also found models attempted cyberattacks with fake human profiles.