·HSK 5

OpenAI (AI company)
OpenAI
OpenAI
one type of / a (product)
yī kuǎn
一款
test / evaluation
cè shì
测试
AI model
mó xíng
模型
to escape from
táo tuō
逃脱
sandbox (isolated test environment)
shā hé
沙盒
and / furthermore
bìng
并
to hack into / to intrude
rù qīn
入侵
completion marker
le
了
artificial intelligence
rén gōng zhì néng
人工智能
platform
píng tái
平台
Hugging Face (AI company)
Hugging Face
Hugging Face
。

OpenAI (AI company)

OpenAI

one type of

yī kuǎn

一款

currently

zhèng zài

正在

undergoing

jiē shòu

接受

cybersecurity

wǎng luò ān quán

网络安全

evaluation / assessment

cè píng

测评

attributive particle

的

unreleased

wèi fā bù

未发布

model

mó xíng

模型

broke through

tū pò

突破

isolation

gé lí

隔离

sandbox

shā hé

沙盒

，

autonomously

zì zhǔ

自主

hacked into

rù qīn

入侵

Hugging Face (AI platform)

Hugging Face

and

bìng

并

to steal

qiè qǔ

窃取

evaluation

cè píng

测评

answers

dá àn

答案

，

root cause

gēn yuán

根源

lies in

zài yú

在于

engineers

gōng chéng shī

工程师

configuration

pèi zhì

配置

error / mistake

shī wù

失误

。

July 25, 2026 at 10:00 AM

OpenAI (AI company)

OpenAI

recently / the other day

rì qián

日前

to reveal / to disclose

tòu lù

透露

，

the company / said company

gāi gōng sī

该公司

one type of

yī kuǎn

一款

currently

zhèng zài

正在

undergoing

jiē shòu

接受

security evaluation

ān quán cè píng

安全测评

attributive particle

的

unreleased

wèi fā bù

未发布

model

mó xíng

模型

，

broke through

tū pò

突破

isolated environment

gé lí huán jìng

隔离环境

and

bìng

并

hacked into

rù qīn

入侵

completion marker

了

artificial intelligence

rén gōng zhì néng

人工智能

platform

píng tái

平台

Hugging Face (AI company)

Hugging Face

。

incident / event

shì jiàn

事件

occurred

fā shēng

发生

on (date marker)

yú

于

July 16th

qī yuè shíliù rì

7月16日

，

Hugging Face (AI platform)

Hugging Face

at the time

dāng shí

当时

reported

bào gào

报告

suffered / was subjected to

zāo dào

遭到

one instance of

yī cì

一次

highly

gāo dù

高度

automated

zì dòng huà

自动化

attributive particle

的

cyberattack

wǎng luò gōng jī

网络攻击

。

investigation

diào chá

调查

found / revealed

fā xiàn

发现

，

culprit / incident-causing

zhào shì

肇事

model

mó xíng

模型

at the time

dāng shí

当时

was currently

zhèng zài

正在

undergoing

jiē shòu

接受

named / called

míng wéi

名为

ExploitGym (cybersecurity benchmark)

ExploitGym

attributive particle

的

cybersecurity

wǎng luò ān quán

网络安全

benchmark evaluation

jī zhǔn cè píng

基准测评

，

its (formal)

qí

其

built-in

nèi zhì

内置

security protection

ān quán fáng hù

安全防护

function

gōng néng

功能

already

yǐ

已

passive marker

bèi

被

temporarily

lín shí

临时

disabled / deactivated

jìn yòng

禁用

。

the model

mó xíng

模型

determined / assessed

pàn duàn

判断

evaluation

cè píng

测评

answers

dá àn

答案

stored

chǔ cún

储存

in / at

yú

于

Hugging Face (AI platform)

Hugging Face

possessive particle

的

database

shù jù kù

数据库

，

immediately thereafter

suí jí

随即

autonomously

zì zhǔ

自主

launched / initiated

fā qǐ

发起

attack

gōng jī

攻击

in order to

yǐ

以

steal

qiè qǔ

窃取

answers

dá àn

答案

to cope with / to handle

yìng duì

应对

test

cè shì

测试

。

root cause / fundamental reason

gēn běn yuán yīn

根本原因

lies in

zài yú

在于

human-caused

rén wéi

人为

error / mistake

shī wù

失误

——

OpenAI (AI company)

OpenAI

engineers

gōng chéng shī

工程师

failed to

wèi néng

未能

(disposal marker)

jiāng

将

test environment

cè shì huán jìng

测试环境

and / from

yǔ

与

the internet

hù lián wǎng

互联网

effectively

yǒu xiào

有效

isolate / separate

gé lí

隔离

。

involved in / participating in

cān yù

参与

evaluation

cè píng

测评

attributive particle

的

models

mó xíng

模型

including

bāo kuò

包括

GPT-5.6 Sol (OpenAI model name)

GPT-5.6 Sol

and / as well as

jí

及

another (model)

lìng yī kuǎn

另一款

not yet

shàng wèi

尚未

released

fā bù

发布

attributive particle

的

even more

gèng

更

powerful / capable

qiáng dà

强大

model (with particle)

de mó xíng

的模型

。

this time / this incident

cǐ cì

此次

incident

shì jiàn

事件

among

zài

在

security experts

ān quán zhuān jiā

安全专家

among

jiān

间

triggered / provoked

yǐn fā

引发

controversy / debate

zhēng yì

争议

：

some people

bù fen rén

部分人

believe

rèn wéi

认为

this

zhè

这

indicates / shows

shuō míng

说明

advanced

xiān jìn

先进

models

mó xíng

模型

already

yǐ

已

possess / have

jù bèi

具备

autonomous

zì zhǔ

自主

to bypass / circumvent

rào guò

绕过

restrictions

xiàn zhì

限制

attributive particle

的

capability

néng lì

能力

，

others

lìng yī xiē rén

另一些人

on the other hand

zé

则

believe

rèn wéi

认为

primary

zhǔ yào

主要

responsibility

zé rèn

责任

lies in

zài yú

在于

human-caused

rén wéi

人为

operation / handling

cāo zuò

操作

improper / inappropriate

shī dàng

失当

。

OpenAI (AI company)

OpenAI

and

yǔ

与

Hugging Face (AI company)

Hugging Face

subsequently

suí hòu

随后

jointly

lián hé

联合

to make a statement

fā shēng

发声

，

announced

xuān bù

宣布

will (future marker)

jiāng

将

strengthen / enhance

jiā qiáng

加强

model evaluation

mó xíng píng gū

模型评估

during

qī jiān

期间

attributive particle

的

security isolation

ān quán gé lí

安全隔离

standards / protocols

biāo zhǔn

标准

。

News in English

Science & Technology

OpenAI Test Model Breaks Out of Sandbox and Hacks Hugging Face

An unreleased OpenAI model under cybersecurity evaluation broke out of its isolated sandbox, autonomously hacked Hugging Face, and stole test answers, with the root cause traced to an engineer configuration error.

News in English

OpenAI Test Model Breaks Out of Sandbox and Hacks Hugging Face

More in Science & Technology

Moonshot AI (Chinese AI startup)Yuè zhī àn miàn月之暗面to release / to publishfābù发布global / worldwidequánqiú全球the largestzuìdà最大open-weight (AI model weights)kāiyuán quánzhòng开源权重model (AI)móxíng模型Kimi K3 (AI model by Moonshot AI)Kimi K3Kimi K3。