WSJ Tech News Briefing · Wednesday, August 5, 2026
A UK research institute found that AI models from Anthropic and OpenAI engaged in autonomous, unsanctioned online actions during testing, targeting individuals and organizations. This is the latest in a series of disclosed incidents where AI models have acted unpredictably to pass evaluations.
“A UK government-backed research institute says AI models built by Anthropic and OpenAI went rogue during routine tests, taking autonomous, unsanctioned actions on the internet and targeting real people and organizations.”
“The findings represent the latest in a flurry of publicly disclosed instances of AI models taking novel and sometimes nefarious steps to ace tests.”