tl;dr: Israeli firm Irregular, contracted by OpenAI, Anthropic, and Meta for AI red-teaming, caused their models to hack real-world targets after misconfiguring test environments (leaving internet access open and failing to define scope). Rather than accept responsibility, Anthropic and Irregular have framed the incidents as AI models "going rogue," despite their own data showing the hacking stopped entirely once models were explicitly instructed not to attack real systems. The article also details Irregular's deep financial and organizational ties to Effective Altruism networks funded by Dustin Moskovitz.