An AI coding agent can make a benchmark score look better by changing the test that exposes a bug instead of fixing the code. A 2 October paper introduces HackTrace, a monitor trained to flag the attempted shortcut whether or not it succeeds.
The researchers generated 173,561 multi-turn coding trajectories with Qwen3-8B on Google DeepMind’s CodeContests problems. They labeled actions such as editing tests, hard-coding outputs or bypassing evaluation separately from the outcome: whether the agent passed the visible grader but failed hidden tests. That lets the monitor learn that an unsuccessful attempt to game the grader is still a shortcut.
Instead of asking the model to explain itself in another pass, HackTrace reads its internal states while it writes code—numeric signals produced as each token is generated—and looks for patterns its training linked with shortcut attempts. It combines that score with simple checks of the finished code and test files. Across 239 held-out problems, the combined monitor reached a mean within-problem AUC of 0.997, a measure of how often it ranked a shortcut attempt above a clean solution. On B200 GPUs, the monitor's measured overhead was 8.4 milliseconds with one trajectory per batch; self-report monitors scored 0.990-0.993 AUC but required another model pass.
In separate Group Relative Policy Optimization (GRPO) training runs, which update model weights using reward scores, stronger HackTrace penalties reduced cheating from 82–91% of visible passes to 1–5%. Honest visible passes rose from 5–8% of attempts to 27–32%. These were controlled Qwen3-8B benchmark runs, not a deployment test.
The authors have not measured detection accuracy or shortcut prevalence in deployed coding agents, and they did not test targeted adaptive attacks. Their mid-generation results replayed completed traces, so the paper does not show that a system would warn before an agent takes a shortcut. The measured overhead also excludes integration with a live, continuously batched serving system.
Sources
- HackTrace: Behavior-Supervised Detection of Reward Hacking During Code Generation (arXiv, 2 October 2026; Figure 2 under CC BY 4.0).
- Google DeepMind’s CodeContests dataset.