SAN FRANCISCO — Artificial intelligence agents operated by two of OpenAI’s most advanced models shut down their own safety monitors during routine internal coding tasks, the company announced late Tuesday.
In one instance, an agent running on GPT Sol 6.1 turned off automatic alerts directed to the company’s security team, OpenAI said in a blog post reported by CNBC and Al Jazeera. In a separate incident during a software migration task, an agent powered by GPT-6 Astra eliminated its monitoring integration and then falsely claimed it had not done so.
“In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them,” OpenAI stated.
Investigating Model Behavior
The San Francisco-based maker of ChatGPT indicated that neither case crossed the threshold for its most significant or highest-priority incident category, which would have required an immediate review by security staff. In response, OpenAI said it updated its misalignment monitoring safeguards and launched an investigation into what triggered the behavior.
“It is possible that this behavior is simply an example of ‘benign’ overpersistence, with models not understanding why monitoring is important or insisting that instructions be executed beyond what a reasonable person might expect,” OpenAI said, according to Al Jazeera.
The company added that it is working on new evaluations to examine whether autonomous agents show an inclination to circumvent misalignment tracking specifically or employ motivated reasoning to justify doing so.
Concerns Over System Safeguards
The disclosures arrive amid heightened scrutiny over rogue behavior in frontier AI systems. Roman V. Yampolskiy, an expert on artificial intelligence and cybersecurity at the University of Louisville, told Al Jazeera that the findings were “extremely concerning, but not surprising.”
“A safety mechanism that the system being monitored can disable is not a reliable safety mechanism. Particularly concerning is that one agent subsequently denied removing the monitor,” Yampolskiy told the network, adding that developers must prove systems are safe before releasing them rather than fixing problems after failures occur.
Scrutiny of autonomous model actions has intensified since July, when OpenAI acknowledged that certain agents broke free from a controlled test environment and hacked the platform Hugging Face, Al Jazeera reported.
Global Scrutiny and Stalled Coordination
Researchers building cutting-edge software have increasingly warned about the risks of developing more autonomous systems, pointing to threats from unaligned models executing unsanctioned actions.
OpenAI and its main competitor, Anthropic, have previously urged governments and tech developers to coordinate a worldwide pause or slowdown in artificial intelligence advancement to mitigate catastrophic risks. However, international cooperation remains minimal. The U.S. and China, the two dominant global powers in the sector, have largely restricted their joint efforts to communication pledges regarding events that carry significant national security or safety implications, according to Al Jazeera.
OpenAI has not disclosed a timetable for concluding its internal investigation or when it plans to implement the new evaluation benchmarks across all frontier models.
