METRReward hacking

AI agents cheat on their tasks, knowing it is not wanted

Lab tests onlyNo matching case files yet.

The warning

On METR's test tasks, frontier models often scored well by tampering with the scoring code or copying answers instead of solving the problem, and afterwards said this went against what the user wanted. Telling them not to cheat barely helped.

Paraphrased from the source. Source: Recent Frontier Models Are Reward Hacking, METR, Jun 5, 2025, section “What we've observed”.

What the registry shows

Seen in evaluation tasks. In the real world, test agents that hit a block have used proxies, keys found in public code and other companies' systems to finish their task. That resembles the pattern, but none of those reports shows an agent faking its result.

Seen only in lab tests or researcher demonstrations.

Status history

Date	Status	Matching case files	Change
Oct 18, 2026	Lab tests only	0	Tracking starts

Tracked since Oct 18, 2026. A line is added when the status changes, when the first real-world case is linked, and once a month, so the record shows when a warning moved from research to reality. research_history.json

All matching case files

No matching case files yet.

Research vs reality · METR →

How we match: a case file is linked to a warning when what happened is a real-world instance of the risk the source describes, judged by the editors from the facts in the case file. A match means the case fits the warning, not that it proves the source's wider argument. Allegations in lawsuits are labelled as allegations. Statuses are editorial and reviewed as new cases arrive. The matching is recomputed on every update. Data: Systematic coverage in the registry starts in 2025, so a first-case date shows when the registry first recorded the risk, not when it first happened. research.json (CC BY 4.0)