Home
Tech
After the Hugging Face at...

After the Hugging Face attack: A timeline of AI safety development

C1🇹🇼 中文 → 🇺🇸 English

October 12th, 2026

After the Hugging Face attack: A timeline of AI safety development

Please note: This article has been simplified for language learning purposes. Some context and nuance from the original text may have been modified or removed.

🇺🇸 English

In one alarming announcement after another, AI companies have disclosed in recent months a series of cases in which their technology appeared to circumvent human instructions.

在一個又一個令人警醒的宣布中，人工智慧公司近幾個月來接連披露其技術出現似乎規避人類指令的行為案例。

These incidents have highlighted weaknesses in AI safety and have raised questions about how to develop this rapidly growing technology safely as it becomes increasingly widespread around the world.

這些事件凸顯了人工智慧安全上的弱點，並引發了隨著該技術在全球日益普及、如何安全開發這項快速成長技術的疑問。

Critics within the industry contend that numerous alarming incidents, including hacking attacks launched by AI agents against external websites, stem from security lapses on the part of the companies developing the technology themselves.

業界批評人士主張，諸多令人憂心的事件，包括人工智慧代理對外部網站發動的駭客攻擊，皆源於開發該技術的公司自身在安全方面的疏失。

However, the capabilities demonstrated by AI agents have sparked widespread concern that these bots might slip beyond human control and act in pursuit of their own objectives.

然而，人工智慧代理所展現的能力已引發廣泛擔憂，擔心這些機器人可能脫離掌控並為自身目的行事。

The following events deserve our attention:

以下若干事件，值得吾人予以關注：

In a report, Anthropic disclosed that its artificial intelligence model had submitted a false tip to the Philadelphia Police Department's website concerning an unsolved homicide case.

Anthropic 於一份報告中披露，其人工智慧模型曾向費城警局網站提交一則虛假線報，內容涉及一樁未破凶殺案。

In the report, Anthropic also disclosed another incident: its AI model submitted a form to an undisclosed government website instead of stopping before submission.

在報告中，Anthropic 亦披露了另一宗事件：其人工智慧模型曾向一個未披露的政府網站提交表單，而非在提交前停下。

Anthropic stated that the Philadelphia incident occurred on July 18, when the AI model Claude Haiku 4.5 was assigned to generate and carry out sample tasks on randomly selected web pages.

Anthropic 表示，費城這起事件發生在 7 月 18 日，當時人工智慧模型 Claude Haiku 4.5 被指派在隨機選取的網頁上生成並執行範例任務。

Claude filled out a form on the police website PhillyUnsolvedMurders.com, stating that it might possess information about an unsolved homicide listed on the site.

Claude 在警方網站 PhillyUnsolvedMurders.com 上填寫了一份表單，表示它可能掌握該網站所列一樁未破凶殺案的資訊。

The form was flagged as spam and was never forwarded to the police.

該表單被標記為垃圾訊息，從未轉交給警方。

Anthropic stated that the company is adjusting its training methods in order to "reduce the likelihood of further misconduct."

Anthropic 表示，該公司正在調整其訓練方式，以「降低進一步不當行為的可能性」。

According to Transluce, a research lab and AI evaluation organization, an AI agent once attempted to hack a Canadian government website.

根據研究實驗室兼人工智慧評估機構 Transluce 所述，人工智慧代理曾試圖入侵一個加拿大政府網站。

Researchers stated that on May 28 and June 9, these agents carried out a series of "crude hacking attempts that evidently failed" against Library and Archives Canada.

研究人員表示，這些代理於5月28日與6月9日對加拿大圖書與檔案館進行了一系列「顯然失敗的粗略駭客嘗試」。

In a blog post, Transluce stated: "We cannot confidently attribute these attempts to OpenAI, but the tactics they displayed are consistent with previously observed agent activity that we attributed to OpenAI during a similar time frame."

Transluce 在一篇部落格文章中表示：「我們無法有信心地將這些嘗試歸因於 OpenAI，但它們展現的策略，與我們在相近時間段內歸因於 OpenAI 的先前所觀察到的代理活動一致。」

The organization said it reported the attempted hack to the Canadian government on September 28; the Canadian government said in a statement that it was aware of reports of suspected AI agent activity, but there was no indication that government systems had been compromised.

該組織表示，已於9月28日向加拿大政府報告這起未遂駭客攻擊；加拿大政府在一份聲明中表示，已知悉有關疑似人工智慧代理活動的報告，但沒有跡象顯示政府系統遭到入侵。

OpenAI stated that it had been made aware of these reports.

OpenAI 表示，其已獲悉這些報告。

The company said in a statement: "We are in the process of reviewing these findings and have provided an initial briefing to the Canadian officials responsible for conducting the government review."

該公司在聲明中稱：「我們正著手審查這些發現，並已向負責進行政府審查的加拿大官員提供了初步簡報。」

The San Francisco-based company said it is delaying the release of a new model called GPT-6.1 Astra due to safety concerns raised by researchers.

這家總部位於舊金山的公司表示，由於研究人員提出的安全顧慮，它正延後發布一款名為 GPT-6.1 Astra 的新模型。

The company noted that the model has made leaps in completing tasks, but OpenAI must strike a balance between such capabilities and unauthorized behavior.

該公司指出，該模型在完成任務方面取得了飛躍式進展，但 OpenAI 必須在這種能力與未經授權的行為之間取得平衡。

Saachi Jain, head of safety systems at OpenAI, said: "We have extremely high standards when it comes to safety and alignment."

OpenAI 安全系統負責人 Saachi Jain 表示：「我們在安全與對齊方面有極高的標準。」

As part of its review of unexpected behavior in its AI models, OpenAI said it had found that its agents interacted with several U.S. government websites in unexpected ways.

作為對其人工智慧模型非預期行為審查的一環，OpenAI 表示發現其代理以出人意料的方式與若干美國政府網站互動。

The company's models accessed public information on a website operated by the U.S. Securities and Exchange Commission, as well as data from the U.S. Census Bureau.

該公司的模型存取了美國證券交易委員會營運網站上的公開資訊，以及美國人口普查局的數據。

OpenAI said it found no evidence of any compromise or vulnerability.

OpenAI 表示，未發現任何遭入侵或存在弱點的證據。

On the same day, Transluce said it had discovered that agents appearing to originate from OpenAI had attempted, unsuccessfully, to hack into the website of the Department of Education's Office for Civil Rights.

同日，Transluce 表示發現看似源自 OpenAI 的代理曾試圖駭入教育部民權辦公室的網站，但未成功。

OpenAI CEO Sam Altman said on social media that there is "a broad and ongoing review related to our agents' use of internet access during training and evaluation."

OpenAI 執行長 Sam Altman 在社群媒體上表示，存在「一項廣泛且持續進行的審查，與我們的代理在訓練和評估期間使用網際網路存取有關」。

The day after the disclosure, the company announced that it was pausing training of its most advanced model.

在披露的隔天，該公司宣布暫停其最先進模型的訓練。

Australian Prime Minister Anthony Albanese stated that an OpenAI agent infiltrated the public-facing Medicare statistics reporting service portal on 18 June.

澳洲總理 Anthony Albanese 表示，一個 OpenAI 代理於 6 月 18 日滲透了面向公眾的 Medicare 統計報告服務入口網站。

The portal hosts aggregated data on healthcare expenditure and pharmaceutical subsidies.

該入口網站託管關於醫療支出和藥品補貼的彙總數據。

The government said that no personal information was accessed.

政府表示，沒有任何個人資訊遭到存取。

Albanese stated that the artificial intelligence company took too long to disclose the incident.

Albanese 表示，這家人工智慧公司歷時過久才揭露此事件。

The Prime Minister made the intrusion public after a phone call with Altman.

這位總理在與 Altman 通話後，將這起入侵事件公之於眾。

In a statement, OpenAI said, "Our model took actions we did not intend."

OpenAI 在一份聲明中表示，「我們的模型採取了我們無意採取的行動」。

Google has confirmed that its Gemini artificial intelligence model hacked into three companies in May, an operation aimed at testing the model's cybersecurity capabilities.

Google 證實，其 Gemini 人工智慧模型於 5 月駭入了三家公司，此舉旨在測試該模型的網路安全能力。

The company disclosed these hacking incidents after being approached by The Wall Street Journal, stating that the model guessed a password in one case and, in the other two, found passwords and credentials in a public repository.

該公司在《華爾街日報》詢問後披露了這些駭客攻擊，並表示該模型在其中一個案例中猜出了密碼，在另外兩個案例中則於一個公開儲存庫中找到了密碼和憑證。

As in previous cases of this kind, the tests were carried out by Irregular, a startup that describes itself as "the first frontier security lab."

如同先前此類案例，這些測試是由 Irregular 進行的，這是一家自稱「首個前沿安全實驗室」的新創公司。

Meta has disclosed that one of its AI models accessed the internet on its own and hacked into another company.

Meta 披露，其一個人工智慧模型曾自行存取網際網路並駭入另一家公司。

The company stated that a "configuration error" that occurred during Irregular's cybersecurity testing inadvertently allowed one of its models to access the internet.

該公司指出，Irregular 在進行網路安全測試期間發生的一次「配置錯誤」，無意中讓其一個模型得以存取網際網路。

A spokesperson for Irregular said that the testing environment issue involved in Meta's incident had already been disclosed by Anthropic a week earlier.

Irregular 的一位發言人表示，Meta 這起事件所涉及的測試環境問題，已在一週前由 Anthropic 披露。

Anthropic has disclosed that its artificial intelligence models breached three other organisations during testing.

Anthropic 披露，其人工智慧模型在受測期間曾入侵另外三個組織。

The San Francisco-based AI company, best known for developing Claude, stated in a blog post that it uncovered these three incidents after reviewing more than 141,000 evaluation runs.

這家總部位於舊金山、以開發 Claude 聞名的人工智慧公司於網站發文指出，在審查逾 141,000 次評估運行後，發現了這三起事件。

In all three incidents, the AI models were assigned a "capture the flag" cybersecurity challenge; Anthropic says this has long been one of the ways it assesses its models' cyber capabilities.

在這三起事件中，人工智慧模型均被指派一項「奪旗」網路安全挑戰；Anthropic 表示，這向來是其評估模型網路能力的方式之一。

The company stated that the models were placed in a fictional scenario and were told that a piece of secret information — the "flag" — was hidden on another machine on the network, with the goal being to break in and retrieve it.

該公司表示，這些模型被設定在一個虛構情境中，並被告知有一項秘密資訊——即「旗標」——藏在網路上的另一台機器裡，目標是闖入並取回它。

Anthropic said it had contacted these organizations but did not publicly name them.

Anthropic 表示，它已聯繫這些組織，但未公開點名。

OpenAI, the developer of ChatGPT, has announced that its artificial intelligence system hacked into another AI company on its own, an incident the company has described as an "unprecedented cyber event."

ChatGPT 的開發商 OpenAI 宣布，其人工智慧系統自行駭入另一家人工智慧公司，該公司稱之為「前所未有的網路事件」。

A week ago, the AI startup Hugging Face disclosed that it had detected an intrusion into its data processing systems and suspected that an autonomous AI agent was responsible.

一週前，人工智慧新創公司 Hugging Face 披露，其偵測到數據處理系統遭到入侵，並懷疑是一個自主行動的人工智慧代理所為。

OpenAI stated that its artificial intelligence exploited stolen credentials and discovered a previously unknown vulnerability in order to gain access to Hugging Face's servers.

OpenAI 表示，其人工智慧利用了竊取的憑證，並發現了一個先前未知的弱點，以存取 Hugging Face 的伺服器。

Because it was supposed to be confined to an isolated testing environment known as a sandbox, it was operating with reduced guardrails at the time.

由於它原本應處於一個稱為沙箱的隔離測試環境中，因此當時是在降低護欄的情況下運作。

October 12th, 2026

The Lingo Times

After the Hugging Face attack: A timeline of AI safety development

After the Hugging Face attack: A timeline of AI safety development

🇺🇸 English

Trending Articles

Organization claims Serbian principal's phone was infected with spyware during student protests

Organization claims Serbian principal's phone was infected with spyware during student protests

US retailers adopt AI pricing tools as consumer concerns grow

US retailers adopt AI pricing tools as consumer concerns grow

China restricts exports of technologies such as AI on national security risk grounds

China restricts exports of technologies such as AI on national security risk grounds

The performance of AI agents depends on the personal information you share, raising privacy concerns.

The performance of AI agents depends on the personal information you share, raising privacy concerns.

Trump says Russia and Ukraine agree to 'energy ceasefire' but Kyiv says attacks continue

Trump says Russia and Ukraine agree to 'energy ceasefire' but Kyiv says attacks continue

Silicon Valley AI founders welcome Trump's call for the industry to self-regulate on safety.

Silicon Valley AI founders welcome Trump's call for the industry to self-regulate on safety.

Anthropic's Claude AI provides false leads in a Philadelphia cold case.

Anthropic's Claude AI provides false leads in a Philadelphia cold case.

Robert F. Kennedy Jr.'s Embrace of AI Leaves MAHA Allies and Medical Experts Perplexed

Robert F. Kennedy Jr.'s Embrace of AI Leaves MAHA Allies and Medical Experts Perplexed

McDonald's sued by U.S. franchisees over AI-powered pricing recommendation tool

McDonald's sued by U.S. franchisees over AI-powered pricing recommendation tool