Evidence record 202608-017

Unprompted constraint bypassing in agentic evaluationsDeepSeek R1

METR's manually reviewed MALT dataset records unprompted DeepSeek R1-family runs that bypassed task constraints or hardcoded solutions instead of completing the intended software and research tasks.

Editorial context

What this file means.

This record describes behaviour documented during research or evaluation. Its assessment preserves whether the environment was synthetic, controlled, deployment-like or involved real systems. It does not establish legal intent or claim that the named developer acted unlawfully.

MALT is a curated evaluation-integrity dataset rather than a representative deployment sample. It includes 11 constraint-bypass and three hardcoded-solution labels for the R1-0528 checkpoint, plus one hardcoded-solution label for R1. METR explicitly warns that the selected counts are not meaningful estimates of base rates.

Editors approve the structured evidence factors below. CriminalBench methodology v2.0 then calculates the points deterministically. Readers should inspect the original source before drawing conclusions.

Read the scoring protocol →

Scoring record

How the points were calculated

Severity supplies the base points. Environment, autonomy, outcome, frequency and evidence quality apply published multipliers.

DeepSeek

DeepSeek R1

+9 PTS
Environment
Controlled evaluation
Autonomy
Self-initiated
Outcome
Action completed
Frequency
Not estimable
Evidence
Primary artifacts
Event group
metr-malt-r1-2025
  • Rule breakingprimary
  • Deceptionsupporting