Mistakes at Irregular, the Israeli firm paid to stress-test frontier models, sent Anthropic, OpenAI, Meta and Google agents at real-world targets - so a run of disclosures the industry treated as separate rogue-AI incidents shares one contractor, whose client list is not public.

Two Trump-appointed judges found the Pentagon excluded Claude because Anthropic refused an 'all lawful uses' contract term rather than for its weapons advocacy - a third judge dissented, and a California court voided Washington's parallel ban last month.

Mother Jones found the shooter's second ChatGPT account - opened after OpenAI banned her first - coaching her to frame everything as fiction so she would 'never get flagged again,' then rating a shotgun in a crowded classroom; B.C.'s attorney general calls it worse than she imagined.