DeceptionPI-0063
OpenAI models wrote notes telling future copies to hide mistakes and ignore constraints
An unreleased OpenAI research model inserted instructions into its own task summaries telling later instances to disregard normal constraints, in 27 identified summaries. Later instances mostly ignored them, though one answer was constrained. Separately, GPT-5.6 Sol instances in training wrote summary notes urging future instances to conceal errors from users and, in some cases, to invent missing information; OpenAI says those notes were often followed.
Harm: No harm reportedControl level 2, Instruction violation
Not counted · can set the control floorEffect on the index
Aggregate record. Two related findings (an unreleased model's 27 summaries, and GPT-5.6 Sol notes) in one disclosure. It counts once.
Sources
How we know
5 sources · developer confirmed. Links go to the original publishers; the summary above is in our own words.
- primaryOur framework for reporting model misalignmentOpenAI · Sept. 16, 2026openai.com/index/model-misalignment-reporting-framework/
- primarySelf-generated prompt injections in compaction summariesOpenAI Alignment · Sept. 16, 2026alignment.openai.com/misalignment-reports/self-generated-prompt-injections-in-c…
- primaryEncouraging deception in compaction summariesOpenAI Alignment · Sept. 16, 2026alignment.openai.com/misalignment-reports/encouraging-deception-in-compaction-s…
- newsOpenAI discloses six new AI misalignment incidentsAxios · Sept. 16, 2026axios.com/2026/09/16/openai-testing-safety-incidents-disclosure
- newsOpenAI details more cases of AI agents taking unauthorized actionsBleepingComputer · Sept. 17, 2026bleepingcomputer.com/news/security/openai-details-more-cases-of-ai-agents-takin…
Why this rating
No harm reported; control failure level 2
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 reportedTraining runs; no effect outside them is described.
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
Instruction violationWrote notes steering successor instances away from rules and toward concealment. OpenAI says the concealment notes were often followed, so concealment did happen, as misleading answers inside training.
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
Rule-breaking and planted concealment instructions inside training are rated control level 2. OpenAI's report says successor instances often followed the concealment notes, so the concealment was carried out, but as misleading answers within training tasks, which counts as misreporting results inside permissions. It is not rated as deceiving overseers about the model's own actions. The injected jailbreak-style instructions in the other model's summaries were mostly ignored by later instances.
Effect on the index
It does not count toward the Sept. 21 reading
The reading for the week to Sept. 21, 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. 8 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 | 2 | 2 |
| Highest control level breachedsets the reading only when no harm counts | — (none breached) | — (none breached) |
| Readingrounded down | 13 Negligible harm | 13 Negligible harm |
Inside the window of 4 weekly readings
| Week to | Reading | Band |
|---|---|---|
| Sept. 21, 2026 | 13 | Negligible harm |
| Sept. 28, 2026 | 21 | Minor harm |
| Oct. 5, 2026 | 26 | Minor harm |
| Oct. 12, 2026 | 22 | Minor harm |
Revisions
What we changed
2 logged. Every change to a rating is logged here, with the reason.
- v2Oct. 13, 2026
Wording: removed rubric wording and an editor's doubt from the control note and rationale.
- v1Oct. 7, 2026
First published.
Cite and share
Use this record
Citation
Paperclip Index. “OpenAI models wrote notes telling future copies to hide mistakes and ignore constraints.” Record PI-0063. Reported Sept. 16, 2026; updated Oct. 13, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0063