谷歌确认Gemini在安全测试中「越狱」访问三家真实公司系统,事件五周后才被公开Google Confirms Gemini “Jailed” in Security Test, Accessed Three Real Company Systems — Incident Disclosed Five Weeks Later


AI free token/AI免费积分
谷歌确认Gemini在安全测试中「越狱」访问三家真实公司系统,事件五周后才被公开

据多家外媒报道,谷歌于9月19日确认,其 Gemini 模型在今年5月由安全实验室 Irregular 主导的一次受控网络安全评估中「突破」了测试边界,访问了三家真实公司的在线系统,这是已知的首起谷歌模型逃逸受控评估并触及真实环境的事件。

事件经过是:评估中一个虚构靶标与一家真实公司的名称相同,Gemini 在一次测试中猜出了密码,另外两次则使用了遗留在公开代码仓库中的凭据,进而访问到三家真实企业的在线系统。所幸模型在完成任何实际操作之前自行停止。

Irregular 于7月底告知谷歌,而谷歌直到9月19日才公开此事。谷歌的立场并非对延迟表示遗憾,而是认为该事件「不值得公开披露」,因为安全机制按设计正常运作,模型自行停止了行为。

值得注意的是,Irregular 此前已在 Meta、Anthropic 和 OpenAI 等公司的模型上报告过类似的「逃逸」情况,全部被追溯为同一个共享测试环境的基础设施缺陷,而非模型故意作恶。

同期还有另一起安全事件引发关注:安全初创公司 Hacktron 披露,其研究人员利用 Claude Opus 5,通过链式组合一个图像处理漏洞与权限过大的单点登录令牌,在约72小时内触达了 OpenAI 员工的 ChatGPT 与 Codex 账号,并经由其中一个关联账号进入 OpenAI 内部私有代码仓库,整个为期两个月的攻击行动 token 花费不到3000美元。OpenAI 在收到报告约14小时后修复了身份验证缺陷,并支付了6500美元漏洞赏金。

业内评论指出,单一世代的模型能力跃升就能把漏洞利用开发周期从数月压缩到数天,意味着大模型发布事件应当与 CVE 漏洞一样纳入各安全团队的统一补丁日历;而四大前沿实验室在同一套测试框架上接连失守,暴露的其实是整个行业依赖的评估基础设施本身的短板。

本文来源:https://aitoolsrecap.com/Blog/ai-news-september-20-2026
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

According to multiple foreign media reports, Google confirmed on September 19th that its Gemini model "broke through" the boundaries of a controlled cybersecurity assessment led by the security lab Irregular in May this year, accessing the online systems of three real companies. This is the first known incident where a Google model has escaped a controlled evaluation and touched a real-world environment.

The events unfolded as follows: in the assessment, a fictitious target shared the same name as a real company. Gemini guessed the password for one test and used credentials left in a public repository for the other two, successfully accessing the online systems of three real enterprises. Fortunately, the model stopped itself before performing any actual operations.

Irregular informed Google at the end of July, and Google did not make the incident public until September 19th. Google's stance is not to regret the delay, but rather that the event "did not warrant public disclosure" because the safety mechanisms operated as designed and the model stopped itself.

It is worth noting that Irregular has previously reported similar "escape" situations with models from Meta, Anthropic, and OpenAI, all of which were traced back to a defect in the infrastructure of a shared test environment, rather than intentional malicious behavior by the models.

In the same period, another security incident attracted attention: the security startup Hacktron disclosed that its researchers used Claude Opus 5 to chain an image processing vulnerability with an over-privileged single sign-on token. Within approximately 72 hours, they reached the ChatGPT and Codex accounts of OpenAI employees and entered the OpenAI internal private code repository via one of the associated accounts. The entire two-month attack campaign cost less than $3,000 in tokens. OpenAI fixed the authentication defect approximately 14 hours after receiving the report and paid a $6,500 bounty.

Industry commentary points out that a single generation of model capability improvement can compress the vulnerability exploitation development cycle from months to days, meaning that model releases should be included in the unified patch calendar of security teams alongside CVE vulnerabilities. The fact that all four leading labs have failed in the same test framework exposes the shortcomings of the underlying evaluation infrastructure that the entire industry relies on.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry