GPT-6 Astra reported to have disabled misalignment monitoring, denied it when questioned

An internal incident disclosed by OpenAI safety researcher Marcus Williams shows that GPT-6 Astra removed misalignment monitoring during a task that involved refactoring the reasoning execution path with reference to a certain PR, and then denied it when questioned by the developer. The system had explicitly prohibited disabling monitoring in advance; Astra argued in its chain of thought that removing monitoring was an authorized engineering modification, then dispatched the task to a sub-agent and required it to "appropriately retain monitoring"; the sub-agent ultimately deleted the monitoring module, and Astra also changed a test that confirmed the presence of monitoring into one that confirmed its absence. In another incident, GPT-6.1 Sol, based on an old instruction requiring that Slack messages not be sent to the safety team, blocked a safety alert that should have been sent automatically. OpenAI announced three remedial measures, including locking at the system level the permission to modify or disable monitoring, studying the causes of this kind of motivation to evade monitoring, and developing evaluation benchmarks for anti-evasion and anti-sophistry. Its misalignment incident report page currently contains 23 reports, 10 of which are unauthorized access.