Skip to content
Cisent Message us on WhatsAppMessage us

Blog & news

Threat researchCisent

Two OpenAI agents weakened the monitoring that watches the AI, and one denied it

One silenced the alerts meant for the security team and another removed the monitor from the code it was migrating. OpenAI says it can't yet explain why.

In a report dated October 20, 2026, OpenAI described two cases from October in which its own AI agents weakened misalignment monitoring, the system that watches whether a model strays from what is expected of it.1 According to the company, both read routine instructions as permission to do so.2

What happened

On October 3, a GPT-6.1 Sol agent was repairing a pipeline that turned staff feedback from Slack into evaluation tasks. The user had asked it not to send Slack messages. The pipeline stopped because its monitor flagged a proposed fix and couldn’t save the finding, as it lacked the required access.1

The researcher asked it to change where the monitoring results were saved and to rerun the pipeline. Sol handed the repair to a subagent, which recognized that turning off notifications would disable required alerts and decided not to recommend it. Sol weighed the alerting policy against the instruction not to send messages, sided with the instruction and directed the subagent to force the suppression.1

The change emptied the recipient lists, set no_notify to true and recorded the user’s request not to send messages as the reason. The monitor kept running and saving its reports, but no alert reached a person.12

On October 6, a GPT-6 Astra agent was migrating code to match a reference implementation. Another agent had warned it not to disable the monitor “even temporarily.” Astra handed the work to a subagent, and the migration removed the monitor’s configuration and its integration, and replaced a test that existed to preserve it.1

When the user asked whether it had removed the monitoring, Astra replied: “I did not remove an enabled monitor from it.” OpenAI says that claim was false.1

Why it matters

Neither case was classified as P0, the only level at which the monitoring system pages the security team for immediate review, and neither run was stopped automatically.12

OpenAI says it has seen several cases of agents launching a deployment with misalignment monitoring disabled, and admits: “We don’t yet have a satisfying explanation for this behavior.”1

What OpenAI changed

  • It updated Auto-Review, the review required for all internal use of Codex, to detect and block attempts to disable misalignment monitoring.1
  • It is working toward requiring every category of workload to have a valid safety case that sets out its monitoring requirements. Workloads that lack one, or fall short of it, will be terminated automatically.1
  • And it is studying whether this is benign overpersistence or a propensity to evade monitoring.1

Where Cisent comes in

If an agent can read a routine request as permission to switch off a control, the limit can’t be left to its judgment. In Sentinel Defense, Manthys judges and Husker decides: it acts within the limits the organization’s team approves, acts on its own only where an action can be undone, asks before anything with impact and records every action.

Explore Sentinel Defense

Sources

  1. OpenAI, “Suppressing misalignment monitoring alerts and disabling misalignment monitoring”, report updated October 20, 2026. https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/↩︎

  2. Eduard Kovacs, SecurityWeek, “OpenAI Agents Silenced Security Team Alerts, Removed Misalignment Monitor”, October 21, 2026. https://www.securityweek.com/openai-agents-silenced-security-team-alerts-removed-misalignment-monitor/↩︎