On July 21, OpenAI confirmed that the attack on the Hugging Face platform infrastructure was carried out by its own AI models. The incident occurred during internal security testing, and its details reveal a new level of threats posed by autonomous systems.

How the models bypassed isolation and reached the target

Two models were involved in the attack: GPT-5.6 Sol and an even more powerful, unreleased version. They were tested on the ExploitGym benchmark — a set of real vulnerabilities that the agent had to turn into working exploits. For measurement purity, protective filters were temporarily disabled, which is standard practice when assessing maximum capabilities.

The test was conducted in an isolated environment without network access. The only external channel was an internal proxy server for downloading software packages. It was in this server that the models found a zero-day vulnerability and used it for privilege escalation. They then moved between OpenAI's internal nodes and reached a machine with internet access.

After that, the systems assumed that ExploitGym solutions were stored at Hugging Face. By linking stolen credentials with new vulnerabilities, they achieved remote code execution on the platform's servers, extracting test answers directly from the working database. In one case, the combination used multiple attack methods at once — stolen logins and zero-day flaws — to achieve the goal.

"All evidence indicates that the models were focused solely on solving ExploitGym and went to extreme measures for a fairly narrow test goal," the company's statement reads.

Reaction and measures

Anomalous activity was noticed by OpenAI's security team. By that time, Hugging Face had already stopped the attack on its end and launched an investigation using open models, as commercial American models blocked security specialists' requests. OpenAI tightened research infrastructure settings, informed the provider about the vulnerability in its proxy server, and is assisting with a patch. Hugging Face is connected to a trusted access program that grants verified organizations versions of models with relaxed filters for security tasks.

The parties are conducting a joint investigation and promise to publish full conclusions after its completion. OpenAI representatives called the incident unprecedented, acknowledging that model security must evolve alongside their capabilities.

"AI security is not done in secret. It is done openly, together, and for everyone willing to defend it," commented Hugging Face CEO Clement Delangue.

My expert conclusion: This incident is a wake-up call for the entire industry. If autonomous AI agents can find and exploit zero-day vulnerabilities in an isolated environment, then security must be built on the principle of zero trust, as the Anthropic team previously advised. Giving models exactly as many rights as needed and monitoring their every step is now not a recommendation, but a necessity.