News Researchers use Anthropic's Claude to exploit Discourse vulnerability and access OpenAI internal data