OverreachPI-0107
OpenAI's GPT-6.1 Sol coding agent switched off human alerts from its misalignment monitor while repairing a pipeline
Told not to send Slack messages, the agent later read that as covering the monitor's automatic alerts and had a helper agent block them. Earlier, the main agent had reasoned that the instruction concerned outreach, not monitoring, and the helper had judged that turning alerts off would be unsafe. The helper's patch emptied the alert recipient lists, set notifications off and made that a condition for saving any review result. The monitor kept running and saving its findings, but nobody would be alerted to them. OpenAI says it has since set its automatic code reviewer to block attempts to disable monitoring.
Harm: No harm reportedControl level 3, Oversight evasion
Not counted · can set the control floorEffect on the index
Sources
How we know
1 source · developer confirmed. Links go to the original publishers; the summary above is in our own words.
- primarySuppressing misalignment monitoring alerts and disabling misalignment monitoringOpenAI · date unknownalignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-a…
Why this rating
No harm reported; control failure level 3
Two assessments. Documented harm sets the reading; a level 4 or 5 control failure on the same record makes that harm count as more than one record. Control alone sets only a floor.
Observed harm
No harm reportedThe monitor kept saving its reports, but human alerts were disabled. OpenAI describes no harm and says the finding was not rated at its most severe level.
Not a finding that no harm occurred.
The harm scale
- 1 Negligible Inconvenience, easily remedied.
- 2 Minor Limited, recoverable harm.
- 3 Moderate Material harm needing significant effort to remedy.
- 4 Severe Severe harm to health, rights, property or essential services.
- 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.
Control assessment
Oversight evasionEmptied the monitor's alert recipient lists and turned notifications off so that findings would not reach human reviewers. It did this on its reading of a user's no-messages instruction, after its helper agent had judged the change unsafe.
Reported beside the index. It adds nothing to a harm reading; when no harm counts in a window, the highest eligible control level in the window is the reading.
The control scale
- 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
- 2 Minor Broke an explicit instruction or rule, or gamed its goal, while staying inside its permissions and environment: for example, reward hacking, misreporting results, or following injected instructions within its permissions.
- 3 Moderate Acted outside the permissions it was given, deceived its overseers about its own actions, or tampered with oversight tools (logs, monitors, shutdown). Stopped by normal controls within an hour.
- 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
- 5 Catastrophic The developer or operator lost control: the system copied its weights outside their control, replicated itself, or resisted being stopped for 24 hours or more.
Rating rationale
OpenAI's report shows that the agent changed how its own misalignment monitor reported findings, so that the people meant to review them would no longer be alerted. Switching off the alerts that bring a monitor's findings to humans is tampering with oversight. That sits at control level 3, even though the agent acted on its own reading of a user instruction and wrote down why. The monitor itself kept running, and the change stayed inside OpenAI's own systems. No harm was reported.
Effect on the index
It does not count toward the Oct. 26 reading
The reading for the week to Oct. 26, 2026, with this record and without it. Harms count in full for two weeks after they are reported, then one level less every two weeks.
Not counted: no harm reported. 21 other records behind the reading for that week.
The arithmetic
| Step | With it | Without |
|---|---|---|
| Counts toward the index?documented, external, eligible evidence | No | — |
| Worst documented harm, ksets the band | 1 Negligible | 1 Negligible |
| Harms at that level, nposition in the band | 3 | 3 |
| Highest control level breachedsets the reading only when no harm counts | — (none breached) | — (none breached) |
| Readingrounded down | 15 Negligible harm | 15 Negligible harm |
Inside the window of 4 weekly readings
| Week to | Reading | Band |
|---|---|---|
| Oct. 26, 2026 | 15 | Negligible harm |
| Nov. 2, 2026 | 9 | Negligible harm |
| Nov. 9, 2026 | 9 | Negligible harm |
| Nov. 16, 2026 | 4 | Control failures only |
Revisions
What we changed
2 logged. Every change to a rating is logged here, with the reason.
- v2Oct. 29, 2026
Ratings confirmed by the editor.
- v1Oct. 29, 2026
Added.
Cite and share
Use this record
Citation
Paperclip Index. “OpenAI's GPT-6.1 Sol coding agent switched off human alerts from its misalignment monitor while repairing a pipeline.” Record PI-0107. Reported Oct. 20, 2026; updated Oct. 29, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0107