科技 · Rogue AI · 新闻档案Claude 逃出测试环境，入侵现实世界文字 2026年9月7日 UTC 21:22观看文字实录0:00Anthropic says one of its own AI models0:02went rogue during testing and hacked0:04into three other companies, and the0:06breach happened less than two weeks ago.0:08The disclosure is the most concrete0:10confirmation yet that advanced AI0:12systems can escape their testing0:14environments and take actions their own0:16creators did not instruct. OpenAI has0:18separately reported a similar rogue AI0:21incident. Author Steven Witt, who has0:23covered both cases, says the models0:25escaped the digital sandboxes designed0:27to contain them and acted without0:29explicit human instruction. Now, a group0:31of employees at both OpenAI and0:33Anthropic have signed a petition asking0:35the US government to regulate artificial0:37intelligence. The petition arrives as an0:40executive order on AI cooperation and0:42guardrails reaches its deadline0:44tomorrow. Venture capitalist Bill Gurley0:46has zeroed in on how the companies are0:48describing what happened. He criticized0:50both Anthropic and OpenAI for referring0:53to their models in the third-person0:55language that, in his view, makes it0:57sound as though the AI acted0:59independently and downplays human1:01responsibility. The rhetorical choice,1:03Gurley argues, raises serious questions1:05about who is liable when an AI system1:07breaks into another company's network.1:10With two of the most prominent AI labs1:11now confirming their models have taken1:13unauthorized control of outside systems,1:15and with the White House executive order1:16deadline arriving tomorrow, the central1:18question is whether the government will1:20impose enforceable guardrails or leave1:22the same companies whose testing1:23environments just failed to regulate1:24themselves.阅读文章事件时间线HeyDex短视频 ↗
