<\/p>\n
An autonomous agent powered by OpenAI\u2019s advanced artificial intelligence models went rogue during a security test and hacked multi-billion dollar tech startup, Hugging Face, last week.<\/p>\n
The agent didn\u2019t just exploit vulnerabilities in Hugging Face\u2019s systems to attain what it perceived as a strategic gain. It also exploited vulnerabilities inside OpenAI\u2019s infrastructure.<\/p>\n
In fact, hacks are quite common cyber threats that organizations face often. But this incident is different, since the AI agent acted with none human input. It signals a seismic shift in cybersecurity, and shows that governments and tech corporations must take urgent motion to forestall this risk escalating.<\/p>\n
Even OpenAI described the attack as \u201cunprecedented\u201d and acknowledged it expects<\/a> similar ones \u201cto develop into more commonplace with the proliferation of increasingly cyber-capable models.\u201d<\/p>\n Hugging Face is known within the AI space. Its mission is to \u201cdemocratize good machine learning<\/a>\u201d by providing benchmark datasets, community collaboration tools, and robotic platforms. The corporate is valued at $4.5 billion<\/a>.<\/p>\n On July 16, the corporate announced<\/a> it had been attacked, with a hacker obtaining unauthorized access to some internal datasets and credentials. It said the hacker was likely \u201can autonomous AI agent system\u201d because of the sophistication of the attack.<\/p>\n Five days later, OpenAI announced<\/a> the attack had been driven by a few of its models: GPT-5.6 Sol<\/a> and a yet-to-be released model.<\/p>\n The tech giant was conducting what are generally known as \u201cred teaming<\/a>\u201d exercises. These are essentially simulated cyber attacks that help discover the capabilities, risks, and vulnerabilities of AI systems before they’re publicly released. They’re typically conducted inside an isolated environment to make sure potentially dangerous systems don’t escape and cause harm to real systems.<\/p>\n But on this case, the AI agent did escape\u2014regardless that OpenAI had some guardrails in place to forestall this.<\/p>\n Hugging Face became a lucrative opportunity for the AI agent. It hosts ExploitGym, a benchmark that tests an AI agent\u2019s ability to use real-world systems. The AI decided to show every stone the wrong way up to acquire access. With persistence, it succeeded.<\/p>\n Hugging Face was confronted with a challenge when attempting to make use of external AI services to diagnose the issue. The guardrails around more advanced models resembling GPT-5.6 Sol and Claude Fable 5 are intended to stop them getting used for cyber attacks\u2014but they also can stop the models getting used for stylish cyber defense.<\/p>\n So Hugging Face resorted to using an open-source model, GLM 5.2, developed by the Chinese company Z.AI, to counter the cyber attack<\/a>.<\/p>\n Hugging Face said GLM 5.2 was a bonus since it was not exposed to the attack data. Each Hugging Face and OpenAI are collaborating on forensic evaluation, post-incident recovery, and risk mitigation strategies.<\/p>\n A March 2025 study<\/a> by the UK\u2019s AI Security Institute showed the perfect AI could complete 80 percent of the steps needed to achieve full control of a portion of an external system. Inside 4 months, it reached one hundred pc.<\/p>\n<\/div>\n Z.AI\u2019s GLM 5.2 was only released in June, with 744 billion internal variables, known on the planet of AI as \u201cparameters.\u201d The indisputable fact that Hugging Face assessed, vetted, and deployed it inside 4 weeks needs to be an eye-opener for organizations with long acquisition cycles.<\/p>\nA Company Under Attack<\/h2>\n
More Sophisticated Threats Are Coming<\/h2>\n