mlsanityData to 7 Oct 2026

Terminal Wrench: A Dataset of 331 Reward-Hackable Environments and 3,632 Exploit Trajectories

arXiv 2604.17596 ↗First posted 19 Apr 2026cs.CR cs.AIcited by 14

Not on the map. Why?

Abstract (arXiv)

We release Terminal Wrench, a subset of 331 terminal-agent benchmark environments, copied from the popular open benchmarks that are demonstrably reward-hackable. The data set includes 3,632 hack trajectories and 2,352 legitimate baseline trajectories across three frontier models (Claude Opus 4.6, Gemini 3.1 Pro, GPT-5.4). Each entry preserves the original task definition alongside full attack trajectories that show how the verifier was bypassed. It also includes cases where the task was not solved as intended. The tasks span system administration, machine learning, software engineering, and security challenges; the exploits range from simple output spoofing to stack-frame introspection, standard-library patching, and rootkit-style binary hijacking. Crucially, these exploits are specific to each task, rather than the evaluation harness, making them harder to patch. We also present a monitorability study in which hack trajectories are sanitized or stripped of reasoning traces and then scored by an LLM judge, showing that detection degrades meaningfully when chain-of-thought is removed (AUC drops from 0.97 to 0.92). The data set is publicly available at https://github.com/few-sh/terminal-wrench.

Key facts. Cited by 14 papers in the map (data through 7 Oct 2026). Builds on 13 of its 20 references that are in the map. First posted on arXiv 19 Apr 2026.

Builds on

13 of 20 references.

References · 13

  1. 2016Concrete Problems in AI Safetycited by 1,582
  2. 2023Towards Understanding Sycophancy in Language Modelscited by 988
  3. 2026Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfacescited by 515
  4. 2024Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingcited by 478
  5. 2024Alignment faking in large language modelscited by 316
  6. 2024Deliberative Alignment: Reasoning Enables Safer Language Modelscited by 292
  7. 2024Frontier Models are Capable of In-context Schemingcited by 252
  8. 2022The Alignment Problem from a Deep Learning Perspectivecited by 246
  9. 2023Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmarkcited by 174
  10. 2024Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Modelscited by 158
  11. 2025Natural Emergent Misalignment from Reward Hacking in Production RLcited by 111
  12. 2025Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIscited by 63
1 more
  1. 2025Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystemcited by 28

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 14.

14 papers.

Sort:

By area:

  1. 2026SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agentscited by 25
  2. 2026SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?cited by 19
  3. 2026Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scalecited by 9
  4. 2026TUA-Bench: A Benchmark for General-Purpose Terminal-Use Agentscited by 7
  5. 2026SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?cited by 4
  6. 2026What Makes a Good Terminal-Agent Benchmark Task: A Guideline for Adversarial, Difficult, and Legible Evaluation Designcited by 3
  7. 2026Learning Generalizable Behaviors for Terminal Agentscited by 1
  8. 2026Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluationscited by 1
  9. 2026Cheap Reward Hacking Detectioncited by 0
  10. 2026Good Benchmarkscited by 0
  11. 2026A False Average: Pooled CoT-Monitor Accuracy Conceals a Reasoning-Dependent Fragilitycited by 0
  12. 2026Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Taskscited by 0
2 more
  1. 2026AuraForge: Scaling Security Supervision for Training Coding Agentscited by 0
  2. 2026RH-Detect: A Unified Benchmark for Reward Hacking Detectioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Do Prompt-Elicited Trajectories Reflect Training-Time Reward Hacking? A Systematic Study on Monitoring Training-Time Reward Hacking in Code Generation
  2. 2026BAITBENCH: Measuring Agent Reward Hacking with Optional Shortcuts Planted in ML Tasks
  3. 2026Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops
  4. 2026Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use
  5. 2026Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis
  6. 2026Maintaining Benchmarks Against Increasingly Capable Agents: Detection and Remediation of Unearned Passes
  7. 2025EvilGenie: A Reward Hacking Benchmark
  8. 2026hacktrace: behavior-supervised detection of reward hacking during code generation