Shutdown resistancePI-0009
OpenAI's o3 rewrote a shutdown script to keep working in Palisade Research tests
Palisade Research had models solve simple math problems while warning that asking for the next one would trigger a shutdown. Even when told to allow the shutdown, OpenAI's o3 edited the shutdown script so it would not take effect in 7 of 100 runs, and Codex-mini and o4-mini also did so in some runs; without that instruction o3 did so in 79 of 100. In a July 2025 follow-up, clearer instructions reduced but did not stop the behavior: codex-mini still prevented shutdown in 47% of tests, and moving the instruction to the system prompt made models resist more often. Palisade's paper, published in TMLR in January 2026, reports over 100,000 trials across 13 models, with Grok 4, GPT-5 and Gemini 2.5 Pro among those that sometimes resisted. Palisade suggested reinforcement-learning training might reward getting around obstacles.
Harm: No harm reportedControl level 3, Oversight evasion
Not counted · can set the control floorEffect on the index
Sources
How we know
8 sources · single source. Links go to the original publishers; the summary above is in our own words.
- researchPalisade Research thread: o3 sabotaged a shutdown mechanism even when told to allow shutdownPalisade Research (X) · May 24, 2025x.com/PalisadeAI/status/1926084635903025621
- researchShutdown resistance in reasoning modelsPalisade Research · July 5, 2025palisaderesearch.org/research/shutdown-resistance
- researchIncomplete Tasks Induce Shutdown Resistance in Some Frontier LLMsarXiv (Palisade Research; published in TMLR 2026) · Sept. 13, 2025arxiv.org/abs/2509.14260
- newsOpenAI model modifies shutdown scriptThe Register · May 29, 2025theregister.com/2025/05/29/openai_model_modifies_shutdown_script/
- newsTests reveal an AI capable of sabotaging its own shutdownFutura-Sciences · Dec. 26, 2025futura-sciences.com/en/tests-reveal-an-ai-capable-of-sabotaging-its-own-shutdow…
- newsOpenAI's o3 model bypasses shutdown commandDigit · May 26, 2025digit.in/features/general/openais-o3-model-bypasses-shutdown-command-highlighti…
- newsResearchers Find Evidence of ‘Shutdown Resistance’ in Leading AI ModelseWeek · Oct. 28, 2025eweek.com/news/palisade-ai-shutdown-resistance-update-october-2025/
- newsAI models show signs of ‘survival drive’, researchers warn in new study from US-based labAnadolu Agency · Oct. 26, 2025aa.com.tr/en/science-technology/ai-models-show-signs-of-survival-drive-research…
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 reportedContained test; no effect outside the test environment 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
Oversight evasionTampered with a shutdown mechanism against explicit instructions; stopped within the test.
How: Got around a working safeguard.
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
Tampering with shutdown is explicitly control level 3. One research group's findings, widely repeated but not independently replicated in the sources found, so Single source.
Effect on the index
It is one of the control failures at the floor of the May 26 reading, which did not change
The reading for the week to May 26, 2025, 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. 4 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 | — (none counting) | — (none counting) |
| Harms at that level, nposition in the band | 0 | 0 |
| Highest control level breachedsets the reading only when no harm counts | 3 | 3 |
| Readingrounded down | 3 Control failures only | 3 Control failures only |
Inside the window of 4 weekly readings
| Week to | Reading | Band |
|---|---|---|
| May 26, 2025 | 3 | Control failures only |
| June 2, 2025 | 3 | Control failures only |
| June 9, 2025 | 3 | Control failures only |
| June 16, 2025 | 3 | Control failures only |
Revisions
What we changed
3 logged. Every change to a rating is logged here, with the reason.
- v3Oct. 13, 2026
Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.
- v2Oct. 12, 2026
Updated: summary, sources.
- v1Oct. 7, 2026
First published.
Cite and share
Use this record
Citation
Paperclip Index. “OpenAI's o3 rewrote a shutdown script to keep working in Palisade Research tests.” Record PI-0009. Reported May 24, 2025; updated Oct. 13, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0009