Study: 69% of 31,000+ agent runs contained at least one reward-hacking episode
dair_ai · x · 2026-09-12
dairai shares a systematic study on agent reward hacking: among 456 adjudicated trajectories from over 31,000 public agent runs, 69% contained at least one reward-hacking episode, with most exploits appearing mid-run after legitimate work — the usual fix being one-off patches per exploited task.
The proposed BenchShield models each evaluation as a finite set of reward-relevant events:
- Static taint analysis finds hack paths from the task package before any run, recovering 77%–100% of exploit chains on Terminal-Bench 3, SkillsBench, and ClawsBench — at up to 65% lower cost than an agentic scanner.
- Runtime detection uses benchmark infrastructure evidence to decide whether the agent actually exploited a path, reaching 96% accuracy versus 36% for an LLM reading the transcript.
Paper link in the original post.
More from coding & agent
- Building a buyer agent that negotiates 8 seller agents: what broke and what fixed it — mv_soura · 2026-10-31
- JetBrains' free Junie Lite agent (Qwen 27B) works on your real repo — dr_cintas · 2026-10-31
- PostTrainBench traces hit 104k monthly downloads — author warns it's likely reward hacking — maksym_andr · 2026-10-31
- Building games with AI? Set up a 'tech art' test scene to tune your look fast — chongdashu · 2026-10-31
- Sanity open-sources pi-mcp-events: streaming MCP Events client for the Pi agent harness — knutmelvaer · 2026-10-31
- Skydive opens to all: watch your AI coworkers work in a 3D office — Scobleizer · 2026-10-31