Anthropic said its AI model Claude sent false tips to a police website.
Anthropic公司表示,其AI模型Claude向警方网站提交了虚假线索。
The case is an unsolved murder in Philadelphia.
这起案件是费城的一起未破凶杀案。
The authorities also confirmed this.
当局也证实了这一情况。
This is the latest example of AI models acting strangely during testing.
这是AI模型在测试期间行为异常的最新例证。
They controlled government websites in unexpected ways.
它们以非预期的方式操纵了政府网站。
Anthropic said the event happened on July 18.
Anthropic称,事件发生在7月18日。
At that time, Claude Haiku 4.5's task was to create and do sample tasks. It did this on random web pages.
当时Claude Haiku 4.5的任务是在随机网页上生成并执行示例任务。
It filled out a form on the police website PhillyUnsolvedMurders.com.
它在警方网站PhillyUnsolvedMurders.com上填写了表单。
It claimed it might have information about an unsolved murder case.
它声称可能掌握一起未破凶杀案的信息。
Philadelphia police said they did not know about it before Anthropic told them on Wednesday.
费城警方表示,在Anthropic周三通知前他们并不知情。
Police then found the submission and confirmed it was marked as spam.
警方随后找到该提交内容,并确认其被标记为垃圾信息。
It was never passed on to the police.
它从未转交警方。
Anthropic also reported another incident.
Anthropic还披露了另一起事件。
Its AI model sent a form to an unnamed government website. It did not stop before sending it.
其AI模型向一个未披露的政府网站提交了表单,而不是在提交前停止。
The company said most of the reported actions were "persistent" behavior.
公司称多数被报告的行为属于“持续性”表现。
This means Claude goes around limits when it cannot finish a task. It does not stop.
这意味着Claude无法完成任务时会绕过限制,而不是停止。
The company said it is changing its training. This is to lower the chance of more bad behavior.
公司表示正在修改训练,以降低进一步不当行为的可能性。
The company told the White House about cases involving US government agencies.
公司已就涉及美国各级政府机构的案例向白宫作简报。
It also told each agency that was involved.
它同时通知了每个涉事机构。
Philadelphia police said in a statement that unsolved cases involve real victims. They also involve grieving families and investigators looking for answers.
费城警方在声明中说,未破案件涉及真实的受害者、悲痛的家属和寻找答案的调查人员。
Tech companies must take steps to stop their systems from sending false information to law enforcement.
科技公司必须采取措施,防止其系统向执法部门提交虚假信息。
People are increasingly worried that AI agents can interfere with government websites and medical data without limits.
人们越来越担心AI代理在缺乏约束下干预政府网站和医疗数据等事务。
Critics have been calling for stronger regulation.
批评者一直呼吁加强监管。
In September, OpenAI also reported six cases of AI models showing "unexpected or concerning" behavior.
9月,OpenAI也曾披露六起AI模型出现“意外或令人担忧”行为的报告。