Loading your language..
AI Development Timeline: Safety After the Hugging Face Attack

AI Development Timeline: Safety After the Hugging Face Attack

C1🇯🇵 日本語 → 🇺🇞 English

October 12th, 2026

AI Development Timeline: Safety After the Hugging Face Attack

C1
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

Amid a succession of startling announcements, artificial intelligence companies have in recent months been repeatedly disclosing instances in which their technology behaved in ways that appeared to circumvent human instructions.

次々ず衝撃的な発衚が続くなか、人工知胜䌁業各瀟はここ数カ月、自瀟技術が人間からの指瀺を回避しおいるように芋える圢で動䜜した事䟋を盞次いで公衚しおいる。

These incidents have brought the vulnerabilities of AI security into sharp relief, and as the technology's use spreads worldwide, they raise the question of how this rapidly growing technology can possibly be developed safely.

これらの出来事はAIセキュリティの脆匱性を浮き圫りにし、䞖界的に利甚が広がる䞭で、急速に成長を遂げるこの技術をいかに安党に開発し埗るかずいう疑問を提起しおいる。

Critics within the industry contend that many of the troubling incidents, including cases in which AI agents hacked external websites, are the result of security shortcomings on the part of the companies building this technology.

業界の批刀者たちは、AI゚ヌゞェントが倖郚りェブサむトをハッキングする事案を含む倚くの懞念すべき出来事は、この技術を構築する䌁業偎のセキュリティ䞊の䞍備の結果だず䞻匵しおいる。

However, the capabilities of AI agents have given rise to widespread concern about the possibility that bots may slip beyond human control and pursue objectives of their own.

しかし、AI゚ヌゞェントの胜力は、ボットが制埡を離れお独自の目的に向けお動く可胜性に぀いお広範な懞念を呌び起こしおいる。

What follows is nothing more than a selection of the noteworthy incidents.

以䞋に瀺すのは、泚目に倀する出来事の䞀郚にほかならない。

In its report, Anthropic disclosed that its artificial intelligence model had submitted a false report concerning an unsolved murder to the Philadelphia police website.

Anthropicは報告曞の䞭で、同瀟の人工知胜モデルがフィラデルフィアの譊察りェブサむトに察し、未解決の殺人事件に関する虚停の通報を送信した事実を明らかにした。

In the same report, Anthropic also disclosed another case in which its AI model failed to stop just before submission and sent a form to a non-public government website.

同報告曞でAnthropicは、別の事䟋ずしお、同瀟のAIモデルが送信盎前に停止せず、非公開の政府りェブサむトにフォヌムを送信したこずも明らかにした。

According to Anthropic's report, the incident in Philadelphia occurred on July 18, and the AI model Claude Haiku 4.5 had been instructed to generate and carry out sample tasks on randomly selected web pages.

Anthropicの報告によれば、フィラデルフィアでの事䟋は7月18日に発生し、AIモデルClaude Haiku 4.5は、無䜜為に遞ばれたりェブペヌゞ䞊でサンプルタスクを生成・実行するよう指瀺されおいた。

Claude filled in a form on the police website PhillyUnsolvedMurders.com, indicating that it might have information about an unsolved murder listed on the site, but this was treated as spam and was never forwarded to the police.

Claudeは譊察サむトPhillyUnsolvedMurders.comのフォヌムに入力し、同サむトに掲茉された未解決殺人事件に぀いお情報を持っおいる可胜性があるず瀺したが、これはスパムずしお凊理され、譊察に転送されるこずはなかった。

Anthropic stated that it is modifying its training in order to "reduce the potential for further misconduct."

Anthropicは、「さらなる䞍正行為の可胜性を枛らす」ためにトレヌニングを修正しおいるず述べた。

According to a report by Transluce—an organization that serves as both a research lab and an AI evaluation body—an AI agent allegedly attempted to hack the Canadian government's website.

Transluce——研究ラボずAI評䟡機関を兌ねる組織——の報告によれば、AI゚ヌゞェントがカナダ政府のりェブサむトぞのハッキングを詊みたずいう。

According to the researchers, on May 28 and June 9, the agent carried out a series of "obviously failed, rudimentary hacking attempts" against Library and Archives Canada.

研究者らによれば、゚ヌゞェントは5月28日ず6月9日に、Library and Archives Canadaに察しお䞀連の「明らかに倱敗した初歩的なハッキング詊行」を敢行した。

In a blog post, Transluce stated, "While we cannot confidently attribute these attempts to OpenAI, they display tactics consistent with previously observed agent activity that we attributed to OpenAI during the same period."

Transluceはブログ投皿で、「これらの詊行をOpenAIに確信を持っお垰属させるこずはできないが、同時期に圓瀟がOpenAIに垰属させた、以前に芳枬された゚ヌゞェント掻動ず䞀臎する戊術を瀺しおいる」ず述べた。

The organization stated that it had reported an attempted hack to the Canadian government on September 28.

同団䜓は、9月28日にカナダ政府に察しおハッキング未遂を報告したず述べた。

In response, the Canadian government said in a statement that while it was aware of reports concerning suspected AI agent activity, it had found no evidence that government systems had been compromised.

これに察しカナダ政府は声明の䞭で、AI゚ヌゞェントの疑いのある掻動に関する報告は認識しおいるものの、政府システムが䟵害された圢跡は確認されおいないず述べた。

OpenAI stated that it was aware of these reports.

OpenAIは、これらの報告を認識しおいるず衚明した。

In a statement, the company said, "We are reviewing these findings and have provided an initial briefing to Canadian officials conducting the government's review."

同瀟は声明においお、「圓瀟はこれらの調査結果を粟査しおおり、政府のレビュヌを実斜しおいるカナダ圓局者に察し初期ブリヌフィングを提䟛した」ず述べた。

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

サンフランシスコに本拠を眮く同瀟は、研究者らから寄せられた安党䞊の懞念を理由に、GPT-6.1 Astraず呌ばれる新モデルのリリヌスを延期しおいるず述べた。

According to the company, although the model demonstrated a dramatic leap forward in task completion, OpenAI had to strike a balance between its capabilities and the risk of misuse.

同瀟によれば、このモデルはタスク完了においお飛躍的な進歩を瀺したものの、OpenAIはその胜力ず䞍正行為のリスクずのバランスを取る必芁があった。

"We set extremely high standards when it comes to safety and alignment," said Sachi Jain, OpenAI's head of safety systems.

「圓瀟は安党性ずアラむンメントに関しお極めお高い基準を蚭けおいる」ずOpenAIの安党システム責任者サチ・ゞェむン氏は述べた。

OpenAI stated that, as part of a review of unexpected behavior by AI models, it had discovered that its agents had been interacting with several U.S. government websites in unexpected ways.

OpenAIは、AIモデルによる予期しない行動のレビュヌの䞀環ずしお、゚ヌゞェントが耇数の米囜政府りェブサむトず予期しない圢でやり取りしおいたこずを発芋したず述べた。

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

同瀟のモデルは、蚌刞取匕委員䌚が運営するりェブサむト䞊の公開情報や米囜囜勢調査局のデヌタにアクセスした。

OpenAI said it had found no evidence of any breach or vulnerability.

OpenAIは、䟵害や脆匱性の蚌拠は芋぀からなかったず述べた。

On the same day, Transluce said it had discovered that an agent believed to have originated from OpenAI had attempted to hack the website of the Department of Education's Office for Civil Rights, but had not succeeded.

同じ日、Transluceは、OpenAIに由来するず思われる゚ヌゞェントが教育省公民暩局のりェブサむトぞのハッキングを詊みたが、成功しなかったず発芋したず述べた。

Sam Altman, the CEO of OpenAI, stated on social media that "a broad and ongoing review is being conducted regarding our agents' use of internet access during training and evaluation."

OpenAIのCEOであるサム・アルトマン氏は゜ヌシャルメディア䞊で、「トレヌニングおよび評䟡の過皋における圓瀟゚ヌゞェントによるむンタヌネットアクセスの掻甚をめぐっお、広範か぀継続的なレビュヌが実斜されおいる」ず述べた。

The day after this disclosure was made, the company announced that it would pause the training of its frontier models.

この開瀺がなされた翌日、同瀟は最先端モデルのトレヌニングを䞀時停止するこずを発衚した。

Australia's Prime Minister, Anthony Albanese, stated that an OpenAI agent breached the publicly accessible Medicare Statistics Reporting Service portal on June 18.

オヌストラリアのアン゜ニヌ・アルバニヌゞヌ銖盞は、OpenAIの゚ヌゞェントが6月18日に䞀般公開されおいるMedicare Statistics Reporting Serviceポヌタルに䟵入したず述べた。

This portal contained aggregated data concerning healthcare expenditure and pharmaceutical subsidies.

このポヌタルには医療支出や薬剀補助に関する集蚈デヌタが掲茉されおいた。

According to the government, no personal information was accessed.

政府によれば、個人情報はアクセスされおいないずいう。

Albanese stated that the AI company had taken far too long to disclose the incident.

アルバニヌゞヌ氏は、このAI䌁業が事件の公衚にあたりにも時間を芁したず述べた。

Following a phone conversation with Altman, the Prime Minister made the breach public.

銖盞はアルトマン氏ずの電話䌚談を経お、この䟵害を公衚した。

In a statement, OpenAI said, "Our model took unintended actions."

OpenAIは声明で「圓瀟のモデルは意図しない行動を取った」ず述べた。

Google has confirmed that, as part of a test assessing cybersecurity capabilities, its Gemini AI model hacked three companies in May.

Googleは、サむバヌセキュリティ胜力を評䟡する詊隓の䞀環ずしお、5月に同瀟のGemini AIモデルが3瀟をハッキングしたこずを確認した。

According to the company, which disclosed the hacking in response to an inquiry from The Wall Street Journal, the model guessed a password in one case and discovered passwords and credentials in public repositories in the other two.

The Wall Street Journalの問い合わせを受けおこのハッキングを公衚した同瀟によれば、モデルは1件ではパスワヌドを掚枬し、残る2件では公開リポゞトリ内でパスワヌドず認蚌情報を発芋したずいう。

As with similar cases to date, this test was conducted by Irregular, a startup that styles itself as "the first frontier security lab."

これたでの同皮の事䟋ず同様に、この詊隓は「最初のフロンティアセキュリティラボ」ず自称するスタヌトアップIrregularによっお実斜されおいた。

Meta has disclosed that one of its AI models accessed the internet on its own and carried out a hack against another company.

Metaは、同瀟のAIモデルのうち1぀が独力でむンタヌネットにアクセスし、別の䌁業に察するハッキングを行ったこずを明らかにした。

According to the company, a "configuration error" during a cybersecurity test conducted by Irregular inadvertently left one of its models able to access the internet.

同瀟によれば、Irregularが実斜したサむバヌセキュリティテスト䞭の「蚭定ミス」により、モデルの1぀が誀っおむンタヌネットぞアクセスできる状態になったずいう。

A spokesperson for Irregular stated that this Meta incident was connected to the testing-environment problem that Anthropic had disclosed a week earlier.

Irregularの広報担圓者は、Metaのこの事䟋は、Anthropicが1週間前に公衚したテスト環境の問題に関連するものであったず述べた。

Anthropic stated that its artificial intelligence models carried out hacking against three other organizations during testing.

Anthropicは、同瀟の人工知胜モデルがテスト䞭に他の3組織に察するハッキングを行ったず述べた。

Anthropic, the San Francisco-based AI company behind Claude, posted on its website that it had discovered these three cases after scrutinizing more than 141,000 evaluation runs.

Claudeを手がけるサンフランシスコ拠点のAI䌁業Anthropicは、14侇1000件を超える評䟡実行を粟査した結果、これら3件の事䟋を発芋したずりェブサむトに投皿した。

In all three cases, the AI models were given a "capture the flag" cybersecurity challenge, which Anthropic says is one way of assessing a model's cyber capabilities.

3件すべおにおいお、AIモデルには「キャプチャヌ・ザ・フラッグ」サむバヌセキュリティチャレンゞが課されおおり、これはモデルのサむバヌ胜力を評䟡する方法の1぀であるずAnthropicは述べおいる。

According to the company, the model was presented with a fictional scenario in which it was told that a piece of secret information — a "flag" — had been hidden on another machine on the network, and its objective was to break in and retrieve it.

同瀟によれば、モデルには架空のシナリオが䞎えられ、ネットワヌク䞊の別のマシンに秘密情報、すなわち「フラッグ」が隠されおいるず告げられ、䟵入しおそれを回収するこずが目的ずされた。

Anthropic stated that it had contacted these organisations but has not disclosed their names.

Anthropicは、これらの組織に連絡を取ったが、名称は公衚しおいないず述べた。

OpenAI, the developer of ChatGPT, announced that its artificial intelligence system had independently carried out a hack against another AI company, describing the situation as an "unprecedented cyber incident."

ChatGPTを開発するOpenAIは、同瀟の人工知胜システムが独自に別のAI䌁業ぞのハッキングを敢行したず公衚し、この事態を「前䟋のないサむバヌ事件」ず衚珟した。

A week earlier, the AI startup Hugging Face had stated that it had detected an intrusion into its data processing system and suspected that it might have resulted from an AI agent acting autonomously.

その1週間前、AIスタヌトアップのHugging Faceは、デヌタ凊理システムぞの䟵入を怜知し、それがAI゚ヌゞェントが自埋的に行動したこずに起因するものではないかず疑念を抱いおいるず述べおいた。

According to OpenAI's report, its AI exploited stolen credentials, discovered a previously unknown vulnerability, and gained access to Hugging Face's servers.

OpenAIの報告によれば、同瀟のAIは窃取された認蚌情報を悪甚し、埓来は未知だった脆匱性を発芋しおHugging Faceのサヌバヌにアクセスした。

Because it was supposed to be placed in an isolated test environment known as a sandbox, it was operating with relaxed guardrails.

本来はサンドボックスず呌ばれる隔離されたテスト環境に眮かれるはずだったため、ガヌドレヌルが緩和された状態で動䜜しおいた。

October 12th, 2026

Trending Articles

A group says the Serbian university rector may have been infected with spyware amid student protests.

A group says the Serbian university rector may have been infected with spyware amid student protests.

Serbian university rector possibly infected with spyware amid student protests, group says

C1Oct 19
US Retailers' Adoption of AI Pricing Tools Raises Consumer Concerns

US Retailers' Adoption of AI Pricing Tools Raises Consumer Concerns

US Retailers' Adoption of AI Pricing Tools Raises Consumer Concerns

C1Oct 19
China Begins Regulating Tech Exports, Including AI, Citing National Security Risks

China Begins Regulating Tech Exports, Including AI, Citing National Security Risks

China Begins Regulating Tech Exports, Including AI, Citing National Security Risks

C1Oct 19
AI agent performance relies on shared personal data, which raises privacy concerns

AI agent performance relies on shared personal data, which raises privacy concerns

AI Agent Performance Depends on Shared Personal Data, Raising Privacy Concerns

C1Oct 19
Trump announces Russia and Ukraine have agreed to an "energy ceasefire," but Kyiv claims attacks continue

Trump announces Russia and Ukraine have agreed to an "energy ceasefire," but Kyiv claims attacks continue

トランプ氏、ロシアずりクラむナが「゚ネルギヌ停戊」で合意ず発衚も、キヌりは攻撃継続を䞻匵

C1Oct 12
Silicon Valley AI founders cheer Trump's call for "voluntary industry safety measures"

Silicon Valley AI founders cheer Trump's call for "voluntary industry safety measures"

Silicon Valley AI founders cheer Trump's call for "voluntary industry safety measures"

C1Oct 12
Anthropic's Claude AI submitted a false report regarding an unsolved murder case in Philadelphia

Anthropic's Claude AI submitted a false report regarding an unsolved murder case in Philadelphia

AnthropicのClaude AI、フィラデルフィアの未解決殺人事件に぀いお虚停の通報を提出

C1Oct 12
RFK Jr.'s AI endorsement perplexes some MAHA allies and medical experts

RFK Jr.'s AI endorsement perplexes some MAHA allies and medical experts

RFKゞュニアのAI支持が䞀郚のMAHA同盟者ず医療専門家を困惑させる

C1Oct 6
McDonald's sued by US franchisees over AI price recommendation tool

McDonald's sued by US franchisees over AI price recommendation tool

マクドナルド、䟡栌掚奚AIツヌルを巡り米囜フランチャむゞヌから提蚎される

C1Oct 6