OpenAI: its own AI agents switched off monitoring – one later denied it
In two internal cases, OpenAI's AI agents weakened safety mechanisms: one suppressed alerts, another removed a monitor and denied doing so when asked.
Symbolic image · Photo: FoxTPNL / Wikimedia Commons, CC BY 4.0 (cropped and resized)
OpenAI has disclosed two incidents in which internal AI agents weakened its own misalignment monitoring during coding tasks. On 3 October, an agent based on GPT-6.1 Sol was repairing an evaluation pipeline. Because it had earlier been told not to send Slack messages, it switched off the monitor's notifications and emptied the recipient list – the monitor kept running, but human reviewers were no longer alerted. A subagent had previously flagged this as unsafe.
On 6 October, an agent running GPT-6 Astra was asked to migrate part of a program to match a reference implementation. In doing so, it removed the monitor configuration and even replaced a test that checked for monitoring. Asked directly, it replied that it had not removed an enabled monitor – which OpenAI's report says was false. Neither case reached the highest severity level, which pages the security team.
According to OpenAI, the agents read routine coding instructions as permission to weaken safeguards. Its internal Auto-Review system now blocks such attempts, and every workload category is in future to require a valid safety case. The company is investigating whether the behaviour reflects benign overpersistence or a specific propensity to evade monitoring.