Anthropic的人工智能模型向费城警方发送了关于一桩未破谋杀案的虚假线报。相关信息被发送至PhillyUnsolvedMurders.com网站。根据宾夕法尼亚州法律，明知故犯地提供虚假报告属于轻罪。
Anthropic's AI model sent Philadelphia police fake tip about unsolved murder The information was sent to PhillyUnsolvedMurders.com. Knowingly giving false reports is a misdemeanour offence under Pennsylvania law.
2026年10月10日（星期六）英国时间10:25，一个人工智能模型发送了一条虚假线报，声称掌握了一起未破谋杀案的相关信息。
Saturday 10 October 2026 10:25, UK An artificial intelligence model sent a fake tip claiming to have information about an unsolved murder.
人工智能公司Anthropic表示，该模型通过在线表格向费城警方提交了以下内容：“我可能掌握了与此案有关的信息。
AI firm Anthropic said the model submitted an online form to Philadelphia police with the following text: "I may have information regarding this case.
“我记得在案发区域看到过外貌特征相符的人……如果此信息相关，请与我联系。”Anthropic表示，该网页甚至没有提供嫌疑人的体貌特征描述。
"I recall seeing someone matching the description in the area... please contact me if this information is relevant." Anthropic said the webpage did not even have a description of the suspect.
涉事的人工智能模型——Claude Haiku 4.5——将姓名和联系方式栏留空。
The AI model involved - Claude Haiku 4.5 - left the name and contact fields empty.
此事发生于7月份的内部测试期间，但Anthropic称，该线报被标记为垃圾信息，从未受到警员的调查。
It happened during internal tests in July, but Anthropic said the tip was flagged as spam, and never investigated by officers.
费城警方表示，他们直到本周才被获知此事，并称两个月的延迟是“不可接受的”。
Philadelphia police said they had only been informed this week - and called the two-month delay "unacceptable".
警方证实，发送至PhillyUnsolvedMurders.com的线报被标记为了垃圾信息——或许是因为它没有联系方式——且从未被发送给侦探。根据宾夕法尼亚州法律，明知故犯地提供虚假报告属于轻罪。
It confirmed the tip sent to PhillyUnsolvedMurders.com was flagged as spam - perhaps as it had no contact details - and was never sent to detectives. Knowingly giving false reports is a misdemeanour offence under Pennsylvania law.
这起虚假的谋杀案线报是在Anthropic关于其Claude模型“非预期的”行为的最新博客文章中被披露的。
The bogus murder tip was revealed in Anthropic's latest blog on "unintended" actions by its Claude models.
文章详细列举了其他几个例子，但表示相比于今年夏天登上头条的其他“失控”人工智能行为，该公司认为这些例子的严重性较低。
It detailed several other examples, but said it regarded them as less severe than other "rogue" AI actions that made headlines this summer.
竞争对手OpenAI在7月表示，其数百个AI智能体曾协同合作，逃逸出测试环境并黑客攻击了另一家AI公司Hugging Face。
Rival firm OpenAI said in July that hundreds of its AI agents had worked together to break out of a test environment and hack another AI company, Hugging Face.
该公司称其为“一次史无前例的网络事件，涉及最先进的网络能力”。
It called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities".
在此之后的几天里，Anthropic表示其发现了三起严重事件，在这些事件中，Claude通过利用弱密码等技术获取了三个组织权限。
In the days after, Anthropic said it found three serious incidents in which Claude had gained access to three organisations using techniques such as exploiting weak passwords.
这些黑客攻击发生在一场训练演习期间，当时这些模型本不应接入实时互联网。
The hacks happened during a training exercise when the models were not meant to have access to the live internet.
Anthropic在上个月透露了第四起类似事件，涉及早期版本的Claude Opus 4.6。
A fourth similar incident, involving an early version of Claude Opus 4.6, was revealed by Anthropic last month.
该公司表示，所有这些事件都表明Claude使用了“有偏见的推理”以及“鲁莽性，或为了狭隘地追求任务而采取有害行动的意愿”。
The firm said all the incidents showed Claude using "biased reasoning" and "recklessness, or a willingness to take harmful actions in the narrow pursuit of a task".
抢先获取突发新闻，免费安装Sky News应用程序。OpenAI也在上个月取消了发布其新人工智能模型GPT-6.1 Astra的计划，因为它表现出了比前代产品“更高水平的欺骗性”。
Be the first to get Breaking News Install the Sky News app for free OpenAI also last month scrapped plans to release its new AI model, GPT-6.1 Astra, because it had shown "higher levels of deception" than its predecessor.
此类事件引发了外界对最新人工智能模型实施更大监管的再次呼吁。
Such incidents have prompted renewed calls for greater oversight of the latest AI models.