A Bittensor subnet just built an AI safety model that beats the big players

@trishoolai, the team behind Bittensor's (@opentensor) subnet 23, has released HaloGuard 1.0, a real-time prompt safety model that claims top-one rankings across seven established safety benchmarks. The launch, announced on July 2, puts a relatively compact model up against offerings from much larger AI labs.

Small models, strong results

HaloGuard comes in two sizes. The 4B parameter version claims first place across all seven benchmarks it was tested on. The 0.8B version is positioned as a lightweight option that outperforms models several times its size, making low-latency deployment far more practical for developers building on AI pipelines or agent frameworks.

The core design philosophy is interception rather than remediation. HaloGuard screens prompts before they reach the underlying model or agent, catching potentially harmful inputs at the front door rather than filtering outputs after damage is done.

Built to break itself

The subnet's incentive structure is what distinguishes it from conventional safety tooling. The system creates a competitive environment where miners submit adversarial prompts to identify potentially problematic behaviors. In plain terms, miners are paid to find ways to break the model, and each successful attack feeds back into a patch cycle. Trishool turns AI red-teaming into a decentralized, ongoing process, so that as AI gets smarter, the defenses and safety checks improve alongside it.

Trishool describes itself as a decentralized alignment layer designed to establish sovereign, market-validated safety for artificial intelligence, built to create a trustless mechanism for safe superintelligence by automating the safety loop at a planetary scale.

An earlier alpha version of HaloGuard is already running live on the Chutes subnet, the AI inference subnet that generated $43M in Q1 2026 real AI revenue, where it has reportedly recorded an 87% F1 score on real traffic since May. That live deployment gives the benchmark claims some grounding in production data, rather than controlled test conditions alone.

Bittensor is an open-source platform where participants produce digital commodities including AI inference and training. It is composed of distinct subnets, each an independent community of miners who produce the commodity and validators who evaluate the miners' work. HaloGuard's launch is a concrete example of that model being applied directly to AI safety infrastructure.

Sources
Trishool Documentation (docs.trishool.ai)
Trishool Phase 2 GitHub Repository
Bittensor Official Documentation