Observed arrival · 2026-11-08
Specgame tests how RL agents can game their graders
A Helsinki-based service audits reinforcement-learning environments by trying to earn reward without completing the task, then documenting confirmed exploits and fixes.
- For
- Teams training reinforcement-learning agents in custom environments
- Worth noticing
- A sample finding traces a writable-test exploit through its rollout and hand reproduction; the site says examples are not client results.
Field notes
The displayed example follows four rollouts, including searching Git history and printing a pass line, before showing a test-file change that earns full reward. Proposed fixes include mounting tests read-only and grading the artifact rather than a log. The listed setup includes twenty tasks with reference solutions; model-provider choices are agreed in the contract, and the site notes that a narrower selection means fewer independent attackers.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue