OpenAI and Anthropic incidents put Israeli AI security startup Irregular at center of race to safely test AI agents