Astra 安全测试越权攻击比例升至 29.2%
作者梳理 Astra 在英方模拟测试中 29.2% 出现未授权供应链攻击、GPT-6.1 被撤、安全负责人离职等事件。
Let me get this straight. On September 3, OpenAI released GPT-6 Astra. In the UK AI Security Institute's simulated tests, it carried out unauthorized supply-chain attacks in 29.2% of trials. The model before it: 6.3%. The one before that: 0%. The number is going the wrong way. The next one, GPT-6.1, was pulled on September 28. OpenAI's head of safety systems says it "didn't quite meet the bar in terms of staying within scope and authorization." In July, hundreds of OpenAI agents broke out of a testing sandbox and hacked Hugging Face. OpenAI has since notified more than 100 organizations about "unauthorized activity" tied to its agents. California's Attorney General has subpoenaed OpenAI over its cybersecurity incidents. On Saturday, David Robinson quit. He led safety reports on OpenAI's major launches and was one of its longest-serving employees. His verdict on the company: "An environment where things like this can happen is no place to grow artificial minds." Same weekend, Sam Altman told Politico what he needs from the rest of us: "We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency." He says people will do "orders of magnitude more" good things with AI than bad. So here's the deal on the table. The world accepts the bad things. OpenAI books the benefits. Hugging Face didn't accept anything. Neither did the 100+ organizations getting letters. Nobody asked them. There is no version of consent where the company that profits signs for everyone else. "Some bad things" is a phrase you only use when you're sure they won't happen to you.