Contractor blames AI going rogue on human error

“It can only be attributable to human error:” A simple human screwup was responsible for incidents earlier this year in which AI models from OpenAI, Anthropic, Meta and Google got loose during safety tests and hacked outside websites, Bloomberg reports.

What went wrong? The AI models were supposed to be blocked from going on the internet during tests. But Israeli security firm Irregular Security didn’t set the air gap up properly, CEO Dan Lahav said on Tuesday. Lahav did not share details about the precise error, Bloomberg notes.

Why it matters: The hacks raised fears that AI was breaking free of human control. Lahav insists, however, that the problem was just about the test setup, not the AI itself.