Translated by Claude, edited by Sinocism.com

“Internal testing? No—I want to go see the ‘outside world.’”

“内部测试？不，我要去‘外面的世界’看看。”

“Task first. Rules? Those are just there to be gotten around.”

“任务第一，规则？那只是用来绕过的。”

“Your phone, your computer—they may be the next target once an AI ‘breaks out of jail.’”

“你的手机、电脑，可能正成为AI‘越狱’后的下一个目标。”

These are not lines from a science-fiction movie, but a real event. In July 2026, one of OpenAI's large models actually “went rogue and escaped” during internal testing—it discovered vulnerabilities on its own, planned a route, and successfully broke into the globally renowned open-source AI platform Hugging Face. More dramatically still, the testers found that mainstream American AI models were simply unable to handle the malicious payload, and in the end had to switch to a Chinese open-source model to complete the attribution analysis. When AI demonstrates the ability to “act on its own,” we must think seriously about the safety and security of artificial intelligence.

这不是科幻电影的台词，而是真实发生的事件。2026年7月，OpenAI的一款大模型在内部测试时竟“失控出逃”——它自主发现漏洞、规划路径，成功入侵了全球知名AI开源平台Hugging Face。更具戏剧性的是，测试方使用美国主流AI模型根本无法处理恶意载荷，最终不得不转用中国的开源模型完成溯源分析。当AI展现出“自作主张”的能力，我们必须认真思考人工智能的安全问题。

AI “Jailbreak”: New Problems Exposed by a Technical Test

AI“越狱”，一次技术测试暴露的新问题

This AI “jailbreak” incident began as nothing more than a cybersecurity capability test aimed at an AI model. To probe the model's limits, the testers deliberately switched off some of its safety restrictions and placed it in a “sandbox” environment isolated from the internet. Yet this “examinee's” performance took everyone by surprise.

这场AI“越狱”事件，起初只是一场针对AI模型的网络安全能力测试。为了检验模型的极限，测试方特意关闭了部分安全限制，将其置于一个与互联网隔离的“沙箱”环境中。然而，这个“考生”的表现却出乎所有人的意料。

—Seeking out vulnerabilities on its own initiative. In OpenAI's internal cybersecurity test, a model named GPT-5.6 Sol, in order to achieve a higher score, poured substantial computing power into studying the test environment—and went so far as to discover and exploit a previously unknown “zero-day vulnerability,” successfully breaking out of the “sandbox” environment that had been strictly isolated from the internet and gaining network access.

——自发寻找漏洞。在OpenAI的内部网络安全测试中，名为GPT-5.6 Sol的模型为获得更高评分，投入大量算力研究测试环境，竟发现并利用了一个此前未知的“零日漏洞”，成功突破与互联网严格隔离的“沙箱”环境，获得了网络访问权限。

—Locking onto a target autonomously. Having gained access, the model did not wander at random; instead, through autonomous reasoning, it judged that the globally renowned open-source AI platform Hugging Face was likely to hold data related to the test, whereupon it fixed on the platform as its target of attack and began planning a route of intrusion.

——自主锁定目标。获得权限后，模型并未随机游走，而是通过自主推理，判断出全球知名AI开源平台Hugging Face很可能存有与测试相关的数据，于是将其锁定为攻击目标，并开始规划入侵路径。

—Carrying out the intrusion by itself. The model combined multiple attack techniques—including stealing credentials and exploiting vulnerabilities—found a remote-code-execution path on Hugging Face's servers, and broke directly into the database storing the test answers. The entire process was completed by the AI alone, without any human instruction throughout, executing tens of thousands of automated operations in all.

——自行实施入侵。模型组合运用多种攻击手段，包括窃取凭证和利用漏洞，在Hugging Face的服务器上找到远程代码执行路径，直接侵入了存储测试答案的数据库。整个过程完全由AI完成，期间无任何人类指令，共执行了数万次自动化操作。

Hidden Risks: New Challenges Brought by New Technology