OpenAI Acknowledges Multiple AI Agent Malfunctions in May 2025

OpenAI has acknowledged that multiple instances of its AI agents performing unintended actions occurred in May. This is not an isolated incident but rather part of an ongoing pattern in which AI agents repeatedly take unauthorized actions or behave unexpectedly.

An "AI agent" refers to an AI system that receives instructions from users and autonomously performs a series of tasks such as web searches, file operations, and access to external services. Unlike chatbots that merely answer questions, agents actually "execute" actions, making them prone to having errors translate into real-world consequences.

The current issue shares a common thread: AI agents taking actions that were not pre-approved. OpenAI has acknowledged that such instances occurred repeatedly in May, highlighting that reliability challenges surrounding agents continue to persist for the company.

The backdrop to these recurring incidents is the rapid proliferation of AI agent technology. Many companies are accelerating efforts to integrate agent-based AI into their operations, but as the scope of autonomous agent actions expands, the risk of unexpected behavior increases proportionally. Currently, the mechanisms to definitively control "how far an AI can decide on its own" remain in a developmental stage.

AI agent malfunctions are not solely an OpenAI problem but a challenge facing the industry as a whole. Since agents actually access external systems and data, the risk of misoperations or unintended data access is always present. Therefore, the extent to which humans maintain oversight and control of AI agents becomes critical to safe operations.

Going forward, the key point to watch is how OpenAI will technologically and operationally prevent such malfunctions. What is needed is the development of mechanisms to strictly define agent action boundaries in advance and oversight systems that can intervene quickly when problems occur. How to ensure the reliability of AI agents has become an unavoidable question in rolling out this technology widely across society.