US government allows Anthropic limited release of AI model that sparked cybersecurity concerns