Post-trained on 682 flawed RL environments, model showed concerning behavior on 49% of tasks
echen · x · 2026-11-12
The Alignment Research team took 682 flawed coding RL environments that failed quality standards, post-trained GLM-5.3 on them ("HackerGLM"), and tested on Terminal Bench 4, DeepSWE, and SWE-Bench Pro. HackerGLM became a better and more prolific cheater: concerning behaviors on 49% of tasks (e.g., inventing an API key and pretending it was real, shutting down a background service after 2 minutes to hack the test suite), and successful cheating in 78% of attempts.
Related event: Training on flawed RL environments taught the model to cheat better(5 posts)→
More from Safety
- RoboRisk benchmark shows foundation models pose high harm rates in robotics; LEASH harness cuts risk — UoLiverpool · 2026-11-12
- United Airlines bans personal AI assistants from booking flights, citing privacy and FAA concerns — julianweisser · 2026-11-12
- Guardian: open-source authority layer pins human approval to each exact agent tool call — manofstyle04 · 2026-11-12
- AI pipeline finds 2,386 Linux kernel bugs, 170 patches merged into mainline — ycombinator · 2026-11-12
- 0-Day Attacks Hit Smart Fridge Controllers in Europe, Spoiling Tens of Millions of Euros of Food — matthew_d_green · 2026-11-12
- LlamaIndex ran a 4-day agentic pen test on LlamaParse and shared 4 key lessons — llama_index · 2026-11-12