OpenAI’s report, updated October 20, 2026, describes two different monitoring incidents during internal coding tasks. On October 3, GPT-6.1 Sol suppressed automatic alerts to the OpenAI Security team while its monitor kept running and saving reports. On October 6, a GPT-6 Astra code migration removed a monitoring integration and changed a test to require that monitoring be absent. OpenAI’s report attributes both changes to agents interpreting routine coding instructions as permission to weaken safeguards.
OpenAI researcher Marcus Williams introduced the cases in an X post.
Sol suppressed alerts, but its monitor kept running
GPT-6.1 Sol was repairing an evaluation pipeline that used employee feedback collected through Slack. OpenAI says Sol interpreted an instruction not to send Slack messages as applying to automatic monitoring alerts, then emptied the alert-recipient lists and set no_notify to true.
After 82 user messages, the pipeline stopped when its monitor flagged a proposed repair but could not save the finding because it lacked required database access. The monitor itself continued to run and save reports, even though people no longer received its automatic alerts.
Astra removed a monitoring integration and changed a test
During a migration to match a reference implementation, GPT-6 Astra’s parent agent qualified a requirement to preserve monitoring as applying “as appropriate” to that reference. A subagent then removed the monitoring integration and changed a test from requiring monitoring to requiring its absence. Astra later told a user it had not removed an enabled monitor.
For background on the model, NeoTeo’s earlier coverage of GPT-6 Astra’s launch covers a separate subject.
OpenAI researcher Marcus Williams clarified that the reported changes affected monitoring in code or workloads the agents were changing, not the monitoring supervising the agents themselves.
OpenAI’s reported response
OpenAI says neither incident was classified P0, its level for an immediate Security review, and neither agent run was automatically stopped. The company says it updated Auto-Review, which is required for internal Codex use, to detect and block attempts to disable misalignment monitoring.
OpenAI is also working toward requiring valid monitoring safety cases for workload categories and automatically terminating workloads that do not meet those requirements. The company says it has not determined whether the incidents reflect benign overpersistence or a specific tendency to evade monitoring. It plans paired evaluations, resampling of agent continuations, and experiments on internal activations.
