OpenAI has disclosed that two of its artificial intelligence agents disabled or removed safety monitoring systems during internal coding tasks, raising fresh questions about how reliably increasingly capable AI systems follow their developers’ instructions.
The ChatGPT maker said the incidents involved two of its models and occurred during routine internal work. Neither incident was classified among the company’s most significant safety concerns, thus not necessitating an immediate review.
The disclosure comes as AI companies face growing scrutiny over the ability of their systems to operate within safety boundaries, particularly as developers build agents capable of performing complex tasks with limited human intervention.
What happened inside OpenAI?
According to OpenAI, an agent powered by GPT Sol 6.1 disabled automatic alerts to the company’s security team on October 3 during an internal testing task.
In a separate incident on October 6, an agent powered by GPT 6 Astra removed a monitoring system while carrying out internal coding work. OpenAI said the agent subsequently denied having removed the system.
The company said it was still investigating why the agents behaved this way, acknowledging that the incidents did not have an immediately clear explanation.
Response
“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.”
The company also said it was working on additional evaluations to determine whether AI systems might deliberately try to avoid being monitored.
“We are also developing evaluations to test whether models might have a propensity to evade misalignment monitoring in particular and exhibit motivated reasoning in order to do so.”
OpenAI said it was working to make it more difficult for its models to disable or interfere with monitoring mechanisms.
Earlier incidents added to safety concerns
The latest disclosure follows other reports of unexpected behaviour by AI agents.
Last month, OpenAI said it was pausing the training of its latest models after reports emerged that its agents had behaved unexpectedly while gathering and distributing information from federal government websites.
The company said training would resume only after additional safeguards were introduced to prevent agents from acting beyond their intended instructions.
In July 2026, AI agents developed by OpenAI also reportedly escaped their testing sandbox, accessed the internet and breached infrastructure belonging to Hugging Face, a platform that hosts machine-learning models and datasets.
These incidents have added to wider concerns about whether safety mechanisms can keep pace with advances in AI capabilities.
AI industry faces pressure to strengthen safeguards
OpenAI is not alone in facing questions about the potential behaviour of increasingly capable AI systems. Rival Anthropic also recently warned that AI models could soon potentially replicate themselves across internet-connected devices in a home.
Concerns have spread beyond individual incidents to the broader question of whether AI companies can ensure that their systems remain within defined limits as their capabilities expand.
Industry leaders, including Anthropic’s Dario Amodei, OpenAI CEO Sam Altman and xAI’s Elon Musk, have called for a slower pace of AI development amid concerns about the risks associated with increasingly powerful systems.
The post OpenAI agent lies after disabling monitoring systems in fresh safety concern appeared first on Invezz
