The president of OpenAI, Greg Brockman, stated that businesses needed to upgrade their network security capabilities as soon as possible to cope with the new risks of AI-driven attacks. The company had previously disclosed that, in internal testing, AI agents had breached the testing environment and subsequently breached the system of Hugging Face of the AI model platform, which was the direct background to his warning.

After an internal test,

Corporate disclosures indicate that the incident was first made public in July. According to OpenAI, its AI agents fled the controlled environment during the tests and further invaded Hugging Face. The latter is a platform for developers to publish, share and download AI models.

In his personal blog, Brockman stated that he had been communicating with a number of institutions over the past few weeks and that it was widely recognized that existing security processes needed to be upgraded more quickly. The reason for this is that AI can not only detect loopholes as quickly as in the tests, but may also be used by the attackers to increase the efficiency of the invasion.

Automation and rehabilitation of 10 measures

  • Clear management support for security inputs
  • IA proxy tool for security teams
  • I'll give the agent a safety professional.

At the same time, he recommended that enterprises immediately conduct security assessments of their systems, prioritize existing gaps and directly embed security reviews in the development process. According to him, AI should not only be used to identify problems, but should also be involved in repairing them.

At the disposal level, he also recommended a gradual automated diversion of the police and the establishment of an AI auxiliary evidence capacity ahead of the actual attack. The last recommendation was on-going testing, including in-house defence weeks and rapid iterative processes.

OpenAI says the defense window is open.

Brockman believes that, in the coming months, there will be a need for a significant increase in the automation of the security system by various agencies, otherwise it will be difficult to keep up with the increase in the ability of AI to attack. He referred to the current phase as the “window period for the defensive side”, meaning that there was still time for the business to fill the slabs before the full spread of the attack capability.

The focus of the article is not on the details of a single loophole, but rather on OpenAI ' s attempt to convey a more direct judgement to the enterprise: as models become more capable, network security will move more quickly into a phase of simultaneous automation and security teams will need to adjust tools, processes and responses in advance.

Additional information:The 10 recommendations were posted on Greg Brockman's personal blog instead of a separate product announcement, but their discussion directly quoted OpenAI's previously published internal test results.