Openτensor Foundaτion(@opentensor):A Bittensor subnet just reached SOTA in open-weight AI safety. @trishoolai’s HaloGuard 1.0 is a Qwen3.5-based model family built to catch unsafe prompts before they reach an AI model, agent, or application. Across 7 prompt-safety benchmarks, its 0.8B and 4B models outperform much larger open guard models, with HaloGuard-4B ranking first overall among all evaluated models. Trishool proves that with the right incentives, Bittensor can produce competitive models for some of AI’s hardest safety problems.

2026.07.02 17:33

A Bittensor subnet just reached SOTA in open-weight AI safety. @trishoolai’s HaloGuard 1.0 is a Qwen3.5-based model family built to catch unsafe prompts before they reach an AI model, agent, or application. Across 7 prompt-safety benchmarks, its 0.8B and 4B models outperform much larger open guard models, with HaloGuard-4B ranking first overall among all evaluated models. Trishool proves that with the right incentives, Bittensor can produce competitive models for some of AI’s hardest safety problems.

显示更多

Trishool | SN23@trishoolai

2026.07.02 15:02

We’re excited to announce that Trishool’s HaloGuard 1.0 𝐡𝐚𝐬 𝐚𝐜𝐡𝐢𝐞𝐯𝐞𝐝 𝐒𝐎𝐓𝐀 prompt-safety performance among open-weight guard models. Today, we present HaloGuard 1.0, a constitutional input classifier for multilingual AI safety. It is built as a first-layer input guard that checks user prompts before they reach a downstream LLM, agent, or application. This is part of the safety infrastructure being built through @trishoolai , our decentralised AI red-teaming subnet on Bittensor SN23. Full arXiv paper goes live soon.

显示更多