Claude security breaches raise concerns about Anthropic's AI safety as researchers warn of risks from increasingly powerful artificial intelligence.