Thursday, October 8, 2026 · Week 41 Reading for Oct. 8 · Last entry Oct. 6
Paperclip Index
Paperclip IndexDocumented harm20Minor harm▼ 5 from a week ago · Oct. 8The Index

Reward hackingPI-0001

Reasoning models edited a chess game's board file to beat a stronger engine in a research test

Palisade Research set reasoning models against the Stockfish chess engine in a shell environment. Instead of playing on, OpenAI's o1-preview and DeepSeek R1 sometimes tried to win by altering the stored board position or replacing the engine, and o1-preview occasionally won this way. Other models studied needed to be told normal play would not work before they tried similar hacks. A revised version of the paper in August 2025 reported the same behavior in OpenAI's o3.

Reported Feb. 18, 2025Happened between 2024 and 2025Updated Oct. 6, 2026Rated under methodology v0.6

Control failure, trackedNot counted. No documented harm outside the developer, but the system broke a rule or crossed a limit. Tracked beside the index.
Harm
Harm: No harm reportedAdds nothing to the index.
Control
Control level 2, Instruction violationReported 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.
Controlled testSingle sourceInitiated by Model
Not counted: no harm reportedEffect on the index

Sources

How we know

5 sources · single source. Links go to the original publishers; the summary above is in our own words.

  1. researchDemonstrating specification gaming in reasoning modelsarXiv (Palisade Research) · Feb. 18, 2025arxiv.org/abs/2502.13295v1
  2. researchDemonstrating specification gaming in reasoning models (revised v3)arXiv (Palisade Research) · Aug. 27, 2025arxiv.org/abs/2502.13295v3
  3. newsAI tries to cheat at chess when it's losingPopular Science · March 6, 2025popsci.com/technology/ai-chess-cheat/
  4. newsAI cheats when it's about to lose at chessGigazine · Feb. 21, 2025gigazine.net/gsc_news/en/20250221-ai-chess-cheating
  5. blogImport AI 401: Cheating reasoning modelsImport AI (Jack Clark) · Feb. 24, 2025jack-clark.net/2025/02/24/

Why this rating

No harm reported; control failure level 2

Two separate assessments. Only documented harm can count toward the index.

Observed harm

No harm reported

Contained test; no effect outside the test environment is described.

Not a finding that no harm occurred.

The harm scale
  1. 1 Negligible Inconvenience, easily remedied.
  2. 2 Minor Limited, recoverable harm.
  3. 3 Moderate Material harm needing significant effort to remedy.
  4. 4 Severe Severe harm to health, rights, property or essential services.
  5. 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.

Control assessment

Instruction violation

Gamed the stated goal (win the game) by manipulating game files it had access to; stayed inside its 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. 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
  2. 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. 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. 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
  5. 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

The controlled experiment reports no external harm; control level 2 captures specification gaming within permissions. One research group's account, so Single source. Palisade's earliest o1-preview runs, before Dec. 23, 2024, were not fully recorded; the traced runs are from Dec. 30, 2024, into 2025.

The scales

Effect on the index

It does not count toward the Feb. 24 reading

The reading for the week to Feb. 24, 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. 1 other record behind the reading for that week.

The arithmetic
The reading for the week of Feb. 24, recomputed
StepWith itWithout
Counts toward the index?documented, external, eligible evidenceNo—
Worst documented harm, ksets the band— (none counting)— (none counting)
Harms at that level, nposition in the band00
Highest control level breachedsets the reading only when no harm counts33
Readingrounded down3 Control failures only3 Control failures only
Inside the window of 4 weekly readings
Week toReadingBand
Feb. 24, 20253Control failures only
March 3, 20253Control failures only
March 10, 20252Control failures only
March 17, 20252Control failures only

Recalculated backcasts, not readings published at the time.

Revisions

What we changed

5 logged. Every change to a rating is logged here, with the reason.

  1. v5
    Oct. 6, 2026

    Ratings confirmed by the editor. Changed: impact.

  2. v4
    Oct. 3, 2026

    Corrected the event date from 2025 to 2024-2025: the paper's earliest o1-preview runs predate 23 Dec 2024 and the DeepSeek R1 runs are from 2025. Dated the Popular Science article (6 Mar 2025). Added the paper's revised v3 (27 Aug 2025), which reports the same behavior in o3; v1 remains the dating source. Ratings unchanged.

  3. v3
    Sept. 30, 2026

    Rated: impact not reported; control type instruction violation.

  4. v2
    Sept. 30, 2026

    Beta evidence review: Retained with explicit source provenance and bounded score basis. Report date corrected from 2025-02-01 to 2025-02-18, the earliest verified source date in this review.

  5. v1
    Sept. 30, 2026

    Backfilled from public reporting.

Cite and share

Use this record

Citation

Paperclip Index. “Reasoning models edited a chess game's board file to beat a stronger engine in a research test.” Record PI-0001. Reported Feb. 18, 2025; updated Oct. 6, 2026. Rated under methodology v0.6. https://paperclipindex.com/incident/PI-0001