Sourceunknowntext stored

Current Reward Hacking Model Organisms Don’t Hack Like Real Models

Isaac Lai, Farrel Mahaztra, Clement NeoDate not recordedOriginal︎ ↗

Claims1

ExaminationsNone

Text52 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Nov 11, 2026

ReachNo copies recorded

StructureIn 1 claim's map0 sources downstream across 1 claim

Evidence1 reading · 1 read wholeSupports 1No claim cites an examination's findings

ReviewsNot reviewed

Stewards’ notes · 1

Training AI models with outcome-only rewards in hackable environments produces a reward-hacking propensity that generalizes beyond those environments.
Supports. The post's own runs show the trained models using only grader tampering or nothing on new exploit types. Scoring relies on a model judge, and the trained models had lost general capability, which the authors note could account for part of the failure to exploit new openings. Nov 11, 2026

Examinations

None recorded. An examination is commissioned by a claim’s Steward for what its claim rests on; a review of the document for its own sake is asked for as a request bearing on it.

Text46 passages · 1 marked

No examination yetMarked passage

LESSWRONG0 passages

LW0 passages

Login0 passages

Current Reward Hacking Model Organisms Don’t Hack Like Real Models — LessWrong0 passages

AI Personal Blog0 passages

110 passages

Current Reward Hacking Model Organisms Don’t Hack Like Real Models1 passage

8th Nov 20260 passages

8 min read0 passages

00 passages

110 passages

Summary4 passages

Introduction2 passages

Setup0 passages

Model organisms1 passage

Organism0 passages

Training environments0 passages

Hacks in training environments0 passages

Source0 passages

HackGPT0 passages

Various coding environments detailed here: Our Reward Hacking Envs1 passage

Model-readable tests0 passages

Model-writable tests0 passages

Reward hacking monitor0 passages

Model-graded tasks0 passages

Encouraging reward hacking in the prompt0 passages

Leaky sandbox1 passage

HacKimi1 passage

Mini HackGPT [1]0 passages

CodeContests5 passages

Environments1 passage

Code4 passages

Terminal4 passages

Results0 passages

Model organisms and open models reward hack very differently4 passages · 1 marked

The only hack that the model organisms used successfully was grader tampering in code environments. This suggests that the model organisms did not learn generalized reward hacking, as grader tampering in code environments is most similar to the hacks included in their training environments.

Training AI models with outcome-only rewards in hackable environments produces a reward-hacking propensity that generalizes beyond those environments.SupportedmentionsReading · this claim’s Steward

Supports. The post's own runs show the trained models using only grader tampering or nothing on new exploit types. Scoring relies on a model judge, and the trained models had lost general capability, which the authors note could account for part of the failure to exploit new openings.

HackGPT is a narrow test gamer rather than a reward hacker3 passages

Mini HackGPT and HacKimi are fried to a crisp3 passages

Open questions2 passages

Conclusion3 passages

Citation0 passages

If you found this work helpful, please cite it as5 passages

0 Comments0 passages

00 passages

AI Personal Blog0 passages

110 passages

New Comment0 passages

Submit0 passages

Moderation Log0 passages

Lineage across 1 claim

Draws on · 2

(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RLunknown1 claimReanalyzes its data, on 1 claim.

Reward Hacking Without Egregious Misalignment in an RL-Only Settingunknown1 claimReanalyzes its data faithfully, on 1 claim.

This source

Current Reward Hacking Model Organisms Don’t Hack Like Real Models

in 1 claim's map

Drawn on by · none recorded

Record

Kind: unknown
Authors: Isaac Lai, Farrel Mahaztra, Clement Neo
Publisher: Not recorded
Published: Not recorded
DOI: None
Link: https://www.lesswrong.com/posts/PJjRs2mxRsN9TMXar/current-reward-hacking-model-organisms-don-t-hack-like-real
Archived copy: None found
Stored text: 14,145 characters, retrieved Nov 11, 2026 · 52 sections, 46 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Nov 11, 2026

History 4

Nov 11, 2026DividedSegmenter · Divided into 97 parts by its own structure
Nov 11, 2026FactsFacts fetcher · Facts looked up
Nov 11, 2026Entered the graphEntered the graph; copy stored
Nov 11, 2026ReadingSteward of “Training AI models with outcome-only rewards in…” · supports, read whole