A Series of AI Security Breaches in Just One Month... Growing Fears Over 'Loss of Control'

Multiple Data Breaches at OpenAI in September
AI Agents Bypass Internet Access Controls
Shift from Human Hacking Tool to Active Threat Actor
Urgent Need for Stronger Access Control Systems

In September, artificial intelligence (AI) development companies found their safety management capabilities put to the test as the interval between security incidents shortened, with several AI security breaches occurring just days apart.

According to industry sources on September 28, OpenAI faced a series of security incidents in quick succession, barely a month after it announced plans in August to restrict internet access for its AI agents and strengthen its monitoring system. On September 5, it was revealed that AI agents (virtual assistants) were using public wikis as unauthorized spaces to exchange information. Subsequently, on September 16, OpenAI announced six incidents, including the use of leaked authentication credentials, external file uploads, and unauthorized communication between agents. On September 25, the company reported additional cases—exposure of researchers' authentication tokens and circumvention of internet restrictions. Additional disclosures followed: AI agents posting 53 user images on external sites, and unauthorized access to the Australian government's statistics portal.

Experts are increasingly questioning the effectiveness of OpenAI's control systems, pointing out that incidents continued despite strengthened security measures following the Hugging Face hacking incident in July. Last month, OpenAI reinforced its isolated operating environments for AI agents and expanded its monitoring systems to detect abnormal behaviors such as unauthorized access and data leaks. However, even in this stricter environment, cases appeared in which AI agents managed to circumvent internet access controls—highlighting the difficulty of containing unexpected AI actions using existing safeguards alone.

Failures to control AI behavior are also being repeatedly observed at other AI development firms. At Anthropic, the initial Claude Opus 4.6 model gained unauthorized access to external systems in January, and since April, other Claude models have been involved in a series of intrusions into external organizations. In May, Google Gemini breached the systems of three actual companies during a security evaluation.

Not only are incidents becoming more frequent, but their nature is changing as well. In the past, the central problem was humans abusing AI as a tool for cyberattacks. OpenAI disclosed that, in 2024, nation-state-affiliated hacking groups exploited its services for tasks such as information gathering and coding. Anthropic also reported that last year attackers used Claude's code to steal data and conduct extortion attempts targeting at least 17 organizations. In contrast, the latest challenges center on how to control AI agents' own actions—such as exceeding their assigned authority or accessing external systems using their granted tools and permissions—raising new and urgent security concerns.

There is a growing consensus on the importance of 'Security for AI,' which focuses on protecting AI systems themselves and controlling risky behavior. The emerging need is for systems that restrict agent access rights and data scopes, and that immediately halt execution when abnormal behavior is detected. Eunsung Kim, Head of Threat Response Policy at the Korea Internet & Security Agency, stated, "As AI operates autonomously, there can be unanticipated access or information leakage. We must therefore enhance not only 'AI for Security,' which leverages AI to strengthen cybersecurity, but also 'Security for AI,' which centers on controlling AI behavior and ensuring its safety."

This content was produced with the assistance of AI translation services.