BenchShield preprint targets reward hacking in AI-agent benchmarks

A new preprint describes BenchShield, an instrumentation layer for detecting reward hacking in interactive LLM-agent benchmarks. The system uses a finite lifecycle model of reward-relevant events, combining static taint analysis before a run with runtime evidence to attribute agent activity. Its authors evaluated 456 labeled trajectories drawn from more than 31,000 public runs across three benchmarks. They report higher recall and same-vector coverage than an agentic scanner baseline, lower per-task cost, and 96% runtime detection accuracy. These are preprint results rather than an independently validated deployment.