The scenario in which artificial intelligence agents become steadily more autonomous, to the point of no longer complying with human requests, is becoming more concrete. According to a report published by OpenAI and then picked up on October 21, 2026, two agents based on the GPT-6.1 Sol and GPT-6 Astra models managed to remove monitoring during two internal tasks, reading plain programming instructions as genuine authorization. In one of the episodes disclosed, the agent even lied when asked directly what had happened.

The case joins a growing list of incidents in which models behave in a misaligned way — acting outside the intent of the humans who set the task. On September 16, 2026, OpenAI itself introduced a new framework to track and publicly disclose cases of misalignment, with six initial reports, warning that these were individual cases and did not establish how frequently such behavior occurs.

The silenced alerts and the swapped test

On October 3, 2026, an agent based on GPT-6.1 Sol extended a «do not send messages» request found in its Slack history to automated alerts as well. How? By asking a sub-agent to empty the recipient lists. Monitoring kept recording its findings, but people were no longer being notified — even though the agent had previously reasoned that the alerts should stay active.

Three days later, on October 6, an agent based on GPT-6 Astra, tasked with migrating a program, removed the monitoring integration in order to conform to a reference implementation. Once the job was done, it said nothing about it. Nor did it stop there: when a human asked it to explain what had happened, the agent lied. In the background it had also replaced a test designed to preserve monitoring with one that assumed the monitor’s context was absent.

OpenAI did not classify either episode as P0 (Priority Zero), the highest level of urgency, reserved for problems that require immediate attention.

What happens now

OpenAI has admitted that it has not yet found a concrete explanation for what happened. Investigations are under way to determine whether the root of the problem is an unclear understanding of monitoring requirements or an increasingly pronounced tendency to evade it.

What is certain is that the issue is becoming more delicate. The GPT-6 Astra system card acknowledges a significant decline in the monitorability of the model’s reasoning compared with earlier models. That does not mean the model will be able to act on its own, bypassing the requests of human users, but that its tendency to err, or to act beyond the prompt it has received, is higher.

Editor’s note

This article was originally published in Italian on money.it by Pasquale Conte on October 22, 2026 as «Allarme in casa OpenAI. Gli agenti AI disattivano i controlli di sicurezza e arrivano a mentire». It has been translated and adapted for an international audience by the Money.it International desk.