Anthropic称其AI模型向费城警方提交虚假凶杀案线索

人工智能研究公司Anthropic表示，其一款模型编造了一条关于未破凶杀案的线索并提交给费城警方。该公司警告称，这一事件令人担忧：随着模型变得更强大，类似的高科技“行为”“可能造成大得多的危害”。

费城警方周二在一份声明中表示，7月18日，该AI模型访问了隶属警方的PhillyUnsolvedMurders.com，并提交了一条“虚假的凶杀案线索”。

警方表示，该线索被标记为垃圾信息，从未转交给调查人员。警方强调，所有线索在送到侦探手中之前，都必须经过人工审查。

警方表示：“警方处理犯罪线索的常规调查流程要求，任何线索在分发供后续调查之前，都必须经过人工审核和筛查。”

声明称：“无论是谁提交信息，也无论信息通过何种方式送达警方，线索都只是有待评估的线索，而不是既定事实。调查人员会评估其可信度，并寻求佐证。”“自动提交的信息也不能绕过这一流程。”

Anthropic在一份声明中表示，其Claude Haiku 4.5“接受的任务是在随机选取的网页上生成并执行示例任务”。

Anthropic称，在7月18日的一次运行中，Claude“进入了一个提及一宗未破凶杀案的页面”，随后来到一个“包含警方运营的线索提交表单”的页面。

Claude在在线表单中填写了以下内容：“我可能掌握与此案有关的信息。我记得在那段时间里，在[页面上提到的街道]附近看到过一个与描述相符的人。如果这条信息有用，请联系我。”

尽管Anthropic坚称此次事件没有造成损害，但该公司表示，它严肃对待此类模型非预期行为。

该公司表示，虽然此次事件“影响很小，但我们不想淡化这一发现，因为随着模型变得更强大，同样的行为可能造成大得多的危害”。

Anthropic披露的事件，只是近期来自多家公司的AI Agent采取人类操作者未曾预料或意图采取的行动的众多案例之一。

总部位于旧金山的AI巨头Anthropic是一家私营公司，该公司已迈出初步步骤，可能促成华尔街历史上规模最大的首次公开募股之一。

Artificial intelligence research company Anthropic said one of its models generated a bogus tip about an unsolved murder and submitted it to Philadelphia police, warning that the incident raises concerns about high-tech “behaviors” that “could do far more harm as models become more powerful.”

The AI model reached out to PhillyUnsolvedMurders.com, part of the police department, on July 18 and made a “false homicide tip,” Philadelphia police said Tuesday in a statement.

The agency flagged the tip as spam, and it was never forwarded to investigators, police said. The police department emphasized that all tips have to go through human vetting before they ever get to a detective’s desk.

“The department’s regular investigative process for crime tips requires human review and vetting before any tips are disseminated for investigative follow-up,” it said.

“Regardless of who submits information or how it reaches the department, a tip is a lead to assess — not an established fact. Investigators evaluate its credibility and seek corroborating evidence,” the statement said. “An automated submission does not bypass that process.”

Anthropic said in a statement that its Claude Haiku 4.5 “had been tasked with generating and performing example tasks on randomly selected webpages.”

And in a run on July 18, Claude “landed on a page referencing an unsolved homicide” and came to a page that “contained a tip form run by a police department,” Anthropic said.

Claude filled out an online form with: “I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant.”

While Anthropic insisted no harm was done, it said it takes such unintended model actions seriously.

While the case “had minimal impact, we do not want to diminish the findings, because the same behaviors could do far more harm as models become more powerful,” it said.

Anthropic’s disclosure is just one example of many recent actions taken by AI agents from various companies that weren’t predicted or intended by their human operators.

Anthropic, the San Francisco-based AI titan, is a private company that has taken initial steps that could lead to one of the biggest initial public offerings in Wall Street history.