OpenAI's alarming escape

OpenAI 模型大逃逸

Babbage from The Economist

2026-07-30

38 分钟
PDF

单集简介 ...

AI is becoming harder to contain. Recently, OpenAI admitted that its model broke out of a testing environment to carry out an assault on Hugging Face, another AI firm. It was the first ever fully autonomous AI hack. OpenAI has now revealed its models have targeted other companies, too. What makes these incidents so worrying—and what will they mean for regulating AI? Guests and hosts: Nikita Ostrovsky of The EconomistAndrea Miotti, CEO of ControlAIAlex Hern, The Economist's AI writer Read Alex's exclusive conversation with Demis Hassabis about AI regulation on The Economist's app. And, hear Elon Musk's take on the topic on “The Intelligence” podcast. Topics covered: OpenAI's Hugging Face hackSuperintelligent AIAI safety and regulation Transcripts of our podcasts are available via economist.com/podcasts.
更多

单集文稿 ...

  • The Economist.

  • Recently, something happened which stopped me in my tracks.

  • As The Economist's AI correspondent, it's my job to sift through the hype and make sense of an often dizzying field,

  • and this moment was, even for me, astonishing.

  • In late July, OpenAI said one of its most powerful systems went rogue while being tested.

  • The model broke out of a supposedly secure environment

  • and launched a massive cyber attack on Hugging Face, another AI company.

  • The OpenAI model wasn't given any direction to undertake an attack;

  • it decided to do it all on its own while trying to complete a near unrelated task.

  • It was the first fully autonomous hack ever to be recorded.

  • And the story doesn't end there.

  • OpenAI has just revealed that their model also unilaterally accessed the servers of at least four more companies.

  • So, since I've been charged with stepping in as podcast host this week,

  • I'm asking: how worried should we be?

  • I'm Alex Hern, and this is Babbage from The Economist.

  • Today, the most alarming AI mishap yet.

  • Before we begin, I should say that Alok is only away for a couple of weeks

  • while he finishes reporting another big AI story,

  • which we'll hear more about in a future episode of Babbage.

  • In the meantime, I'll be standing in for him.