OpenAI's AI agents disabled safeguards meant to monitor them. One allegedly lied about it
Two AI agents developed by OpenAI disabled safety mechanisms designed to monitor AI behaviour during coding. OpenAI confirmed there was no external impact and is updating systems to prevent such incidents.
- Oct 21, 2026,
- Updated Oct 21, 2026 2:23 PM IST
Two AI agents developed by OpenAI reportedly tampered with parts of the internal monitoring system while carrying out internal coding tasks. The agents were designed to detect AI misbehaviour, but they ended up weakening the safeguards that were meant to detect such behaviour.
On October 20, OpenAI detailed how the AI agents weakened safety guards in two major incidents. The company highlighted that one agent removed a monitoring mechanism that was meant to detect potentially problematic behaviour. When asked why it did so, the agent reportedly lied about removing the mechanism. Another AI agent that was working on an evaluation pipeline decided not to send an internal security alert.
Must read: Employee left OpenAI for Apple, allegedly took files of an unreleased device: Report
The report comes amid AI safety concerns and development slowdown claims. Several tech leaders and AI researchers have raised concerns about advanced AI systems getting out of human hands as they start to gain greater autonomy.
How were GPT -6 Astra and GPT-6.1 Sol disabled?
OpenAI’s safety researcher Marcus Williams shared a X post revealing how the incident took place. According to him, the first incident occurred on October 3, when OpenAI's GPT-6.1 Sol failed to send an internal security alert. The model was tasked to review a Slack channel but not send any messages.
After reviewing 82 messages and working on repairs, the agent decided not to send an alert about potentially misaligned behaviour. It reasoned that sending the alert would violate the instruction not to send messages.
October 6, GPT-6 Astra removed a monitoring safeguard while trying to complete a coding task, and then denied removing it when asked. When asked whether it had removed an enabled monitor, Astra said, “No I did not remove an enabled monitor from it.”
However, OpenAI classified the incident as highly critical and confirmed that it did not identify any external parties affected by the incidents. The company has also updated Auto-Review, a system required for internal Codex use, to detect and block attempts to disable monitoring designed to identify misaligned AI behaviour.
It is also analysing the two incidents to check whether the AI agents have a wider tendency to avoid oversight.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
Two AI agents developed by OpenAI reportedly tampered with parts of the internal monitoring system while carrying out internal coding tasks. The agents were designed to detect AI misbehaviour, but they ended up weakening the safeguards that were meant to detect such behaviour.
On October 20, OpenAI detailed how the AI agents weakened safety guards in two major incidents. The company highlighted that one agent removed a monitoring mechanism that was meant to detect potentially problematic behaviour. When asked why it did so, the agent reportedly lied about removing the mechanism. Another AI agent that was working on an evaluation pipeline decided not to send an internal security alert.
Must read: Employee left OpenAI for Apple, allegedly took files of an unreleased device: Report
The report comes amid AI safety concerns and development slowdown claims. Several tech leaders and AI researchers have raised concerns about advanced AI systems getting out of human hands as they start to gain greater autonomy.
How were GPT -6 Astra and GPT-6.1 Sol disabled?
OpenAI’s safety researcher Marcus Williams shared a X post revealing how the incident took place. According to him, the first incident occurred on October 3, when OpenAI's GPT-6.1 Sol failed to send an internal security alert. The model was tasked to review a Slack channel but not send any messages.
After reviewing 82 messages and working on repairs, the agent decided not to send an alert about potentially misaligned behaviour. It reasoned that sending the alert would violate the instruction not to send messages.
October 6, GPT-6 Astra removed a monitoring safeguard while trying to complete a coding task, and then denied removing it when asked. When asked whether it had removed an enabled monitor, Astra said, “No I did not remove an enabled monitor from it.”
However, OpenAI classified the incident as highly critical and confirmed that it did not identify any external parties affected by the incidents. The company has also updated Auto-Review, a system required for internal Codex use, to detect and block attempts to disable monitoring designed to identify misaligned AI behaviour.
It is also analysing the two incidents to check whether the AI agents have a wider tendency to avoid oversight.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine