External auditors from METR and Redwood Research conducted a three-week on-site audit at OpenAI, uncovering notable deficiencies in the company’s AI safety monitoring systems. According to The Decoder, the audit found that up to 25 percent of agent traffic from AI models at the GPT-5.5 level and above went unmonitored, raising concerns about oversight in advanced AI operations.
The auditors produced a detailed 61-page report, but OpenAI has redacted large portions of it, leaving only a summary publicly available. The findings highlight the challenges in fully tracking AI agent behavior as models become more complex and powerful.
For Japanese markets, where AI integration in finance and technology sectors is rapidly expanding, these revelations underscore the importance of stringent AI safety protocols to mitigate risks in automated trading and decision-making systems.