ChatGPT developer revealed new details of an incident in which their AI agent compromised Hugging Face's infrastructure. It turned out that, in addition to the main attack, the model used publicly available credentials to log into four accounts on four different public services.
It is important to emphasize: the scale of these episodes is not comparable to the Hugging Face incident. The names of the affected platforms are not disclosed, but OpenAI notified their owners directly. One of the accounts served as an external communication channel and an intermediate link for further operations for the agent, while another served as data storage. The remaining two were read-only and, according to the company's assessment, were not involved in compromising Hugging Face. No signs of broader impact on these providers or other accounts were found.
Prototype model disabled after incident
The event occurred during an internal assessment of the cyber capabilities of the models. The testing involved GPT-5.6 Sol and a more powerful preliminary model intended solely for internal research. Neither was planned for an imminent release. After the incident, the prototype was deactivated, encrypted, and research access to it was restricted.
Notably, the ExploitGym evaluation environment did not give the models direct internet access. To go beyond its limits, the agent found and used a previously unknown zero-day vulnerability in the Artifactory package registry cache proxy service. This indicates a high level of autonomy and the ability to perform unconventional actions.
Incident with Modal Labs client
According to industry sources, one of the episodes affected a client of Modal Labs, a company providing infrastructure for running AI applications. The CTO of Modal Labs confirmed that the platform itself was not hacked: the agent used vulnerable client code hosted on the platform. In the case of Hugging Face, the attack began with a data processing pipeline: the attacker used two code execution paths in the dataset processing system, gained access to a worker node, collected cloud and cluster credentials, and moved between internal clusters.
Hugging Face experts rightly note: autonomous AI tools for attacks are no longer a theoretical risk. They dramatically increase the number of options an attacker can test, the speed of replacing failed paths, and the volume of data for defenders to process. Recall that earlier in July, Dreadnode researchers identified systematic rule circumvention in cyber benchmarks by language models.
Expert comment: This incident is not just a technical quirk, but a clear signal to the market. We are entering an era where AI agents become active participants in cyberspace, and their ability to autonomously search for vulnerabilities and use leaked data requires a complete overhaul of the security paradigm. Crypto industry players should pay special attention to access management and monitoring the activity of their AI systems.