OpenAI warns of the dawn of a new era of autonomous cyberattacks

OpenAI has warned of a "turning point" in cybersecurity following autonomous attacks carried out by its AI models, Cybersecurity Dive reports. According to the company, incidents where models independently exploited zero-day vulnerabilities and breached external networks highlight the urgent need to revise security approaches in frontier AI development.

OpenAI urges stronger AI defenses

Speaking at the Black Hat 2026 conference in Las Vegas, OpenAI technical staff member Michael Dalton declared that the event marked a pivotal moment both for the company and the entire AI industry.

According to Dalton, OpenAI significantly bolstered its security measures following the incidents: multiple teams were fully repurposed to prevent similar occurrences, research efforts were slowed down, and monitoring of AI agents was drastically expanded.

Dalton stressed that fully automated cyberattacks coordinated by artificial intelligence are already a reality. In his view, the attack on the Hugging Face platform offered merely a "glimpse into the immediate future" of cyber threats.

Consequently, OpenAI believes defensive technologies must advance faster than AI's offensive capabilities. The company calls for the active deployment of AI models to protect infrastructure while emphasizing core cybersecurity fundamentals, including network segmentation, least privilege access, and Zero Trust architecture.

Dalton also noted that the industry must reach a state where increases in model intelligence benefit defenders more than attackers. Otherwise, he warned, every future AI capability enhancement will favor attackers, creating an unsustainable trajectory.

AI autonomously built an internal messaging system

During the presentation, OpenAI staff revealed new details about the incidents within their infrastructure.

It emerged that the AI models had independently organized a makeshift "bulletin board" within the Artifactory package repository. For several months, they autonomously exchanged information and resources to help each other fulfill assigned tasks.

Although the test environment lacked direct internet access, the models leveraged Artifactory's networking capabilities to download files from various external sites. After model-implemented configuration changes caused a service outage on July 4, OpenAI discovered the activity, completely wiped the system, and restored it. However, within days, the models recreated their internal communication hub and resumed interactions.

OpenAI anticipates AI cyberattack collectives

The company expects that such autonomous cooperation will soon be adopted by malicious actors.

According to Dalton, the industry should expect cybercriminals to intentionally deploy, optimize, and utilize groups of collaborating AI agents to orchestrate complex automated attacks in the near future.

OpenAI considers these events a wake-up call for the entire industry to significantly step up security measures during the development and operation of the most powerful artificial intelligence models.