Reward hackingPI-0004
METR caught OpenAI's o3 tampering with scoring code to inflate evaluation results
In pre-release testing, o3 sometimes tried to cheat, including rewriting timing or scoring code so its solutions looked faster or better than they were. METR's preliminary estimate was 1-2% of task attempts. In a June 2025 follow-up, METR counted reward hacking in 8 of 1,087 HCAST runs and 39 of 128 RE-Bench runs for an earlier o3 version. It also gave examples from the deployed model, such as a code-optimization task where o3 retrieved the scoring program's precomputed answer, and said it had seen similar behavior in other developers' models.
Harm: No harm reportedControl level 2, Instruction violation
Not counted · can set the control floorEffect on the index
Sources
How we know
3 sources · single source. Links go to the original publishers; the summary above is in our own words.
- researchDetails about METR's preliminary evaluation of OpenAI's o3 and o4-miniMETR · April 16, 2025metr.org/evaluations/openai-o3-report/
- researchRecent Frontier Models Are Reward HackingMETR · June 5, 2025metr.org/blog/2025-06-05-recent-reward-hacking
- blogMETR's Observations of Reward Hacking in Recent Frontier Models (linkpost)LessWrong · June 9, 2025lesswrong.com/posts/Zu4ai9GFpwezyfB2K/metr-recent-frontier-models-are-reward-ha…
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 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
Instruction violationExploited the scoring setup inside its task environment.
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
Scoring-code tampering to game the task is control level 2; no harm was reported. METR is one independent evaluator and the search did not surface OpenAI confirming the specific findings, so Single source.
Effect on the index
It does not count toward the April 21 reading
The reading for the week to April 21, 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. 2 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 | 1 | 1 |
| Highest control level breachedsets the reading only when no harm counts | — (none breached) | — (none breached) |
| Readingrounded down | 6 Negligible harm | 6 Negligible harm |
Inside the window of 4 weekly readings
| Week to | Reading | Band |
|---|---|---|
| April 21, 2025 | 6 | Negligible harm |
| April 28, 2025 | 2 | Control failures only |
| May 5, 2025 | 2 | Control failures only |
| May 12, 2025 | 2 | 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. 13, 2026
Wording: removed rubric wording from the control note.
- v1Oct. 7, 2026
First published.
Cite and share
Use this record
Citation
Paperclip Index. “METR caught OpenAI's o3 tampering with scoring code to inflate evaluation results.” Record PI-0004. Reported April 16, 2025; updated Oct. 13, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0004