Irregular became the common link in AI security incidents involving OpenAI, Anthropic and Meta. Here’s what actually went wrong.