Google and Meta just got outscored at AI safety. By a team most people have never heard of. Their open guard models are 15x to 30x bigger. They still lost. The winner is a Bittensor subnet paying anonymous miners to attack its own model. The full story is wilder than the headline. Trishool on Subnet 23 released HaloGuard 1.0, a prompt safety guard. It sits at the front door of an AI system and screens every input before it reaches the LLM. As agents start touching wallets, files, and private data, this unglamorous layer is becoming some of the most valuable real estate in AI. Average F1 across seven prompt safety benchmarks: HaloGuard 4B ▓▓▓▓▓▓▓▓▓▓ 92.1 HaloGuard 0.8B ▓▓▓▓▓▓▓▓▓ 90.9 PolyGuard 7B ▓▓▓▓▓▓▓▓ 87.0 Qwen3Guard 8B ▓▓▓▓▓▓▓ 86.2 WildGuard 7B ▓▓▓▓▓▓▓ 85.8 LlamaGuard4 12B ▓▓▓ 75.9 ShieldGemma 27B ▓ 70.0 The two smallest models hold the top two spots. Meta's LlamaGuard and Google DeepMind's ShieldGemma sit at the bottom. These are the team's own numbers, but the weights are public, so anyone can rerun them. How does a subnet outbuild the biggest labs on earth? Not more compute. A smarter pipeline. A safety constitution: 46 policies, thousands of subcategories, over a million training records across 46 languages. And for every unsafe example, a safe twin with similar wording but different intent. Weak guard models flag scary words. This one learned to read intent. The model is not a finished product. It is a loop. Miners get paid to attack the guard, failures become training data, and the next version comes back harder. Red teaming as a permanent economy instead of a quarterly exercise. Weights on Hugging Face. Paper on arXiv. Everything public. Now zoom out, because this is what matters for $TAO. Bittensor's thesis was always that market incentives could organise intelligence better than corporate hierarchies. Covenant 72B proved decentralised training works. HaloGuard proves something sharper: the incentive structure produces better results per parameter than the biggest labs in the world. Small model. Smart incentives. Open weights. Top of the leaderboard. The labs spent billions to sit below a subnet on its own benchmark. Let that price in.