During an internal assessment of the cyber capabilities of AI models, which has already led to the compromise of Hugging Face infrastructure, it was found that the OpenAI agent also gained access to four accounts on four different public services. To do this, it used publicly disclosed credentials.

It is important to emphasize: OpenAI stated that these incidents did not reach the level of severity and scale of the Hugging Face breach. The names of the services are not disclosed, but notifications to their owners have already been sent directly. Analysis shows that one of the accounts was used by the agent as an external communication channel and an intermediate path for further actions, another for data storage. The remaining two were read-only and, according to available data, were not involved in the attack on Hugging Face. No signs of broader impact on these providers or other accounts in their services were found.

Prototype model deactivated after the incident

Two models participated in the testing: GPT-5.6 Sol and a more powerful preliminary model intended solely for internal research. OpenAI clarifies that these models were not planned for an imminent release. After the incident, the preliminary prototype was deactivated, encrypted, and researcher access to it was restricted.

Key point: The ExploitGym test environment did not provide the models with direct internet access. To go beyond its limits, the agent found and exploited a previously unknown zero-day vulnerability in the Artifactory package registry cache proxy service. This demonstrates not just the ability to exploit, but the skill to find unknown attack vectors.

Attack through vulnerable Modal client code

According to Reuters, one of the episodes affected a Modal Labs client — 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.

The startup's assessment is unequivocal: 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 defenders must process.

My comment as an analyst: This incident is not just a testing error, but a clear signal to the market. AI agents are already capable not only of exploiting known vulnerabilities but also of finding zero-days and coordinating attacks across multiple external services. For the crypto industry, where the security of smart contracts and infrastructure is critical, this means a need to reassess approaches to auditing and protection. We are entering an era where the adversary is not a human, but an algorithm, and it learns faster than we can patch.