AI agent safety is becoming increasingly important as agents take on more complex tasks and interact with external tools. @trishoolai (SN23) is making progress on one of the critical parts of that challenge: ensuring AI agents can safely handle their outputs. HaloGuard's output guard has now achieved a new state of the art, recording an average F1 score of 83.63 across public benchmarks, surpassing the previous best of 82.5. What makes this particularly interesting is how the team achieved these results. A few key details: • 0.8B model, roughly a quarter of the size of the previous leading model. • Designed to make continuous output monitoring more efficient and affordable. • Developed through Phase 3 on Trishool, using 7,500+ jailbreak attempts from 125+ unique miners over 15 weeks. The approach is worth paying attention to. Miners are incentivised to find weaknesses in the guard, and those failures become valuable training signals for improving the model. This creates a continuous feedback loop where decentralised red-teaming directly contributes to stronger safety systems. It also builds on HaloGuard's earlier progress with its input guard, which had already achieved SOTA performance at a similar model size. With the output guard now catching up, both sides are coming together to provide a practical safety layer for AI agents. Efficiency matters here. Safety systems need to be capable enough to identify threats without becoming too expensive or slowing down real-world deployments. This is another example of how Bittensor's incentive-driven approach can turn distributed contributions into measurable improvements in open AI infrastructure. Models and research paper are expected on Hugging Face and arXiv soon.