mlsanityData to 14 Oct 2026

Scalable Weak-to-Strong Verifier Hardening for Agent Environments

arXiv 2610.15924 ↗First posted 11 Oct 2026cs.CR cs.AI cs.LGcited by 0

On the map: LLM agents and tool use → Agent threats, guardrails and governance

Abstract (arXiv)

Agent environments are essential for large language model (LLM) development. They use programmatic verifiers to turn complex outcomes into rewards. Verifiers can be attacked into awarding full credit to cheating agents without completing the intended task, causing inaccurate evaluation and misalignment in training. As agent environments scale in numbers and difficulty, so must the repairing of exploitable verifiers. To this end, we ask whether today's model with scaled compute can automatically harden agent environments to prevent potentially stronger future models from reward hacking. We introduce a scalable, fully automatic approach to weak-to-strong verifier hardening. It first sets up an isolated verifier sandbox that agents cannot directly modify, then uses a hacker-fixer loop to find and repair remaining verifier vulnerabilities. We apply this loop to Terminal Bench, KernelBench, and SETA, substantially reducing attack success across these agent environments. The iterative design scales with the hardening budget, producing progressively stronger defenses as the number of loop iterations increases and outperforming the static, non-iterative baseline. An earlier version of our hacker-fixer loop has been adopted by the official Terminal Bench repository as a way to audit exploits in its environments. Project page: https://verifier-hardening.github.io/.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use. Builds on 22 of its 41 references that are in the map. First posted on arXiv 11 Oct 2026.

Builds on

22 of 41 references.

References · 22

  1. 2026Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfacescited by 585
  2. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 467
  3. 2018AI safety via debatecited by 400
  4. 2025KernelBench: Can LLMs Write Efficient GPU Kernels?cited by 225
  5. 2024Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Modelscited by 178
  6. 2024Prover-Verifier Games improve legibility of LLM outputscited by 85
  7. 2025Kevin: Multi-Turn RL for Generating CUDA Kernelscited by 78
  8. 2025School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMscited by 66
  9. 2025ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Casescited by 65
  10. 2025Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimizationcited by 56
  11. 2026CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generationcited by 47
  12. 2026Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Usecited by 31
10 more
  1. 2025Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystemcited by 30
  2. 2025EvilGenie: A Reward Hacking Benchmarkcited by 28
  3. 2026SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agentscited by 28
  4. 2025Mutation-Guided LLM-based Test Generation at Metacited by 25
  5. 2026Detecting Safety Violations Across Many Agent Tracescited by 14
  6. 2026SETA: Scaling Environments for Terminal Agentscited by 12
  7. 2025GAR: Generative Adversarial Reinforcement Learning for Formal Theorem Provingcited by 6
  8. 2026RvB: Automating AI System Hardening via Iterative Red-Blue Gamescited by 3
  9. 2026Hodoscope: Unsupervised Monitoring for AI Misbehaviorscited by 3
  10. 2026Proactive Hardening of LLM Defenses with HASTEcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops
  2. 2026AgentLeak: Cloning Stronger LLM Agent Capabilities onto Weaker Agents Beyond Skill Stealing
  3. 2026Scaling Agentic Verifier for Competitive Coding
  4. 2026AgentVisor: Defending LLM Agents Against Prompt Injection via Semantic Virtualization
  5. 2026Towards Reliable Local Security Agents: Verifiable Post-Training for Linux Privilege Escalation
  6. 2025SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
  7. 2026BackdoorAgent: A Unified Framework for Backdoor Attacks on LLM-based Agents
  8. 2026Verifier-Grounded Agentic RL for Provably Correct Code