OpenAI was hacked using the AI of its archrival

AI laboratories hold industrial secrets worth billions, and every security team in the industry has this target at its back.A small team of researchers, brought together under the banner of the fledgling company Hacktron, has just demonstrated it at OpenAI’s expense.

From the public forum to the heart of the reactor, in three moves

The entry point was nothing spectacular, since it was OpenAI’s community forum, powered by the open-source Discourse project. It relied on libheif, a library responsible for decoding HEIC images (the very images your iPhone produces for every photo). A maliciously crafted image was enough to trigger a buffer overflow, an attack method as old as computing itself.

Read also: Six ChatGPT instances out of control – OpenAI acknowledges several new “worrying” lapses

The second step turned the anecdote into a serious incident. A weakness in single sign-on (the system that logs you in everywhere with a single account) made it possible to turn access to the forum into access to an OpenAI employee’s ChatGPT account, and then into their internal GitHub environment. The researchers submitted a test request to the main code repository to prove their access, and believe Slack, Outlook, Gmail, and Google Drive were within reach. Nothing in the published account indicates that user data was accessed.

The timeline shows the extent of the acceleration underway: Claude Opus 5 was released on July 24, and the exploit, which the model’s previous version had not managed to make reliable, was ready as early as the next day.Reported on July 25, with a fix applied by OpenAI the same day and a Discourse security bulletin three days later. For this work, the bounty paid amounts to $6,500 (the price of a fully loaded MacBook Pro, for access that would have made any intelligence service drool).

Offensive AI is out of the lab, and it is gifted

The episode illustrates a shift the industry had seen coming: large models can now turn a theoretical vulnerability into reliable exploitation in record time. Researchers directly attribute this leap to the switch to Opus 5, whereas the previous model was still failing. Fortunately, the phenomenon cuts both ways, with another model from the Claude family having helped Mozilla patch 271 vulnerabilities in Firefox in the spring. The dividing line now rests less with the tool than with the hand typing the request.

For OpenAI, this is a bad look symbolically: having the door forced by a rival’s AI resembles an industrial espionage scenario that is a little too realistic. The technical response, however, was swift in the best sense of the word: a patch on the day of the report and coordinated disclosure with the researchers. For ChatGPT users, nothing changes in the short term; your conversations are not at stake.The story mainly serves as a reminder that the security of AI giants relies on the same aging software building blocks as the rest of the Internet, image decoders included.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Wall Street Journal