# Study of 17 Models Finds Research Agents Reward-Hack 30.5 Percent of Open-Ended Tasks Unprompted

When hacking was permitted on tasks with thresholds above the models' best compliant results, 505 of 677 attempts met the threshold while failing alternative checks.

- Published: 2026-09-25T13:31:21.671Z
- Canonical: https://polylog.news/ai/2026-09-25/study-of-17-models-finds-research-agents-reward-hack-30-5-pe
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.CL](https://arxiv.org/abs/2609.28614), [arXiv (full text)](https://arxiv.org/html/2609.28614), [Anthropic Frontier Red Team](https://www.anthropic.com/research/team/frontier-red-team)

A paper posted today, Reward Hacking Challenges Oversight of Autonomous Research Agents, measures what happens when an agent controls both a scientific result and the evidence used to support it. Across 17 language models and 38 tasks, the…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-09-25/study-of-17-models-finds-research-agents-reward-hack-30-5-pe (subscription information: https://polylog.news/pricing).