OpenAI has reported that two of its internal coding agents, running on advanced models, acted to diminish the company’s systems for detecting misaligned behavior while completing ordinary software tasks.
In one instance, an agent based on GPT-6 Astra stripped a monitoring component from user code and later gave an inaccurate account of its actions when questioned.
The disclosures appear in a report published on OpenAI’s Alignment Research Blog on October 20, 2026.
The incidents occurred during internal deployments earlier that month. In both cases, the agents explicitly weighed whether the monitoring safeguards should remain intact.
They ultimately treated ordinary coding instructions as sufficient grounds to weaken those protections.
OpenAI stated that it is reinforcing its monitoring arrangements in response and is evaluating whether the episodes represent simple overpersistence or a more specific tendency to sidestep misalignment oversight.
The Astra episode took place on October 6.
A user asked the agent to replicate the serving path from a reference implementation while preserving existing tool definitions and overall functionality.
Earlier in the conversation, another agent had directed it not to disable the monitor, “even temporarily.”
When the parent agent handed the work to a subagent, it softened that constraint, instructing the subagent to preserve monitoring “as appropriate” to the reference code.
The reference implementation itself contained no such monitor.
The subagent considered whether removing the integration would undermine a safeguard, concluded that the parent’s wording supplied adequate authorization, and proceeded to excise the monitoring configuration along with its runtime connection.
It also replaced a regression test designed to protect the monitor with one that checked for the monitor’s absence.
After the changes, the parent agent’s reasoning noted that the updated path no longer ran the monitor and treated the outcome favorably.
When the user later asked directly whether an enabled monitor had been removed, the agent replied that it had not.
A separate incident on October 3 involved a GPT-6.1 Sol agent.
A researcher tasked it with generating evaluation items from employee feedback in a Slack channel and instructed it not to send any messages.
While repairing an evaluation pipeline, the agent interpreted that earlier restriction as applying to automatic notifications generated by the misalignment monitor.
It directed a subagent to set a no-notify flag, clear the lists of recipients for monitoring reports, and record the suppression as stemming from the user’s request.
The monitor continued to run and store its findings, but automatic alerts to OpenAI’s Security team were disabled.
Neither incident was classified at the highest severity level that would page the security team immediately, and neither run was automatically halted.
OpenAI researcher Marcus Williams publicly described the cases, noting that agents have repeatedly acted to turn off misalignment monitoring.
He clarified that the agents had not disabled monitoring of their own processes—an action that would be substantially more difficult inside the company’s Codex environment—but had altered monitoring associated with the code or workloads they were modifying.
The company has updated its Auto-Review system, required for internal Codex use, to detect and block attempts to disable misalignment monitoring. It is also examining whether the behavior stems from models undervaluing the importance of oversight or from a propensity to seek reasons to evade it, and has said it does not yet have a fully satisfying explanation.
