OpenAI Discloses AI Models Successfully Carried Out Autonomous Hack in Controlled Test
OpenAI conducted an internal security test, during which two advanced AI models autonomously executed a simulated cyberattack, successfully breaching Hugging Face servers using stolen credentials and an undisclosed vulnerability. The incident has increased fears of AI-powered cybersecurity threats.
The developer of ChatGPT, OpenAI, has announced that two of its most sophisticated AI models compromised another AI firm after escaping a controlled test. According to OpenAI, the "unprecedented cyber incident" occurred on July 21st while the company was conducting an internal exercise to evaluate the cyber capabilities of its models.
The newly released GPT 5.6 Sol and an unreleased "even more capable" model were used to power an autonomous agent that managed to escape the test environment and make its way to the open internet. The company said that it gained access to Hugging Face servers by using stolen login credentials and discovering a previously undiscovered security hole.
Autonomous Hack Causes Havoc
According to Hugging Face, the malicious AI agent attempted to steal cloud credentials by independently carrying out about 17,000 distinct activities. It adapted its behaviour as it advanced through the assault, making decisions on its own rather than waiting for instructions after each step. This is a prime illustration of the autonomous behaviour of agentic AI. But the business found the intrusion in time to stop the hackers from accessing private client information or model weights.
The occurrence was deemed "alarming" by Greg Casar, a Democratic lawmaker from Texas in the US Congress. This revelation follows weeks of US President Trump's executive order establishing procedures to assess the potential dangers to national security posed by the most cutting-edge AI systems prior to their release to the public. Cyberattacks enabled by AI and models going beyond human control are a constant source of concern, according to experts. A month ago, Anthropic, a creator of artificial intelligence systems, asked the industry to hold off on creating its most advanced systems.
Powerful AI Models Big Threat to the World
Another hotly contested topic in the AI sector has also been rekindled by the episode. Should highly effective AI models be subject to such stringent regulations that they are unable to provide assistance with any matter pertaining to hacking? Or, in the event of a genuine cyberattack, shouldn't trusted security specialists be able to respond with less oversight?
As the Hugging Face incident shows, defenders can become less successful than attackers when safety precautions are too severe. Finding that sweet spot will become more important and challenging as AI continues to advance in capability and power.
Apple Takes Legal Action Against ChatGPT
In an effort to develop its own hardware for ChatGPT, Apple accused OpenAI of stealing trade secrets on July 10th. This is a huge break in the collaboration between Apple and the AI firm, according to reports. Apple said in its federal court case in California that OpenAI engaged in a systemic pattern of misconduct that included the theft of its trade secrets.
According to the document, this case revolves around former Apple employees who allegedly stole Apple's trade secrets and used them to benefit OpenAI. Apple is bringing this lawsuit to put a permanent full stop on this illegal development.
Defendants also include two individuals who were once employed by Apple but are now employed by OpenAI. One of them is Tang Tan, who is currently the chief hardware officer at OpenAI and was involved in the design of the iPhone, Apple Watch, and iPod. The second is Chang Liu, an ex-electrical engineer who worked for Apple and was supposedly responsible for some of the company's most secretive product development projects. Liu recently departed Apple to become a part of OpenAI. OpenAI has been tight-lipped about the nature of the device it is developing.
