Xiaomi built a dedicated hack agent to stress-test LLM training environments until no exploit remained
alejandroll10 · x · 2026-09-23
Blogger bycloudai highlights that Xiaomi, rather than chasing viral headlines, built a dedicated "hack agent" to probe training environments as hard as possible before letting LLMs train in them, iterating until the agent could no longer find a successful exploit in any environment.
The author questions why other frontier labs don't do the same — recklessness or viral marketing? He also notes this is open-model SoTA: models finding exploits is definitely possible; the question is whether you "let" them escape.
More from Models
- JevBench hits HN frontpage, critics allege Jev-class model is a thin Qwen wrapper — airesearch12 · 2026-09-23
- Opus 5.5 One-Shots a Full Prince of Persia Level With Graphics, NPCs and Music — iannuttall · 2026-09-23
- Opus 5.5 Cut Prices 40% and Within a Day It Was Porting C to Rust: 8 Use Cases — alex_verem · 2026-09-23
- Beff Jezos jokes Opus 5.5 is 'post-slop', freeing readers from AI sludge prose — beffjezos · 2026-09-23
- Opus 5.5 One-Shots a Full Prince of Persia Level with NPCs, Sound and Music — iannuttall · 2026-09-23
- GPT-6 Sol Underperforms GPT-5.6 Max on DeepSWE, 68.8% vs 72.7% — banaxi-tech · 2026-09-23