You don’t have to be publicly traded or an IPO-bound frontier lab to push the frontier. State of the art is being built on Bittensor. Trishool's HaloGuard Output Guard for AI safety and alignment is now #1# on the benchmark, outperforming much larger models from @nvidia, @Google, @alibaba_cloud, and @Meta. $TAO

Bittensor's AI economy is starting to show real commercial traction. A growing number of subnets on @opentensor are turning machine intelligence into products people actually pay for. From GPU compute to consumer apps, security and sales tools, these are the subnets leading it. Layer 1: GPU compute and AI inference Compute is still Bittensor's strongest commercial vertical. - @lium_io (SN51) rents out GPU capacity to developers and AI teams. - @engyai (SN53) runs verifiable inference across a distributed miner network. - @chutes_ai (SN64) has earned roughly $3.6M in inference revenue across Q1 to Q3, though Q3 came in lower than the first two quarters. - @TargonCompute (SN4), built by @manifoldlabs, has bought back and burned about 1,772 TAO from revenue and burned another 6,678 TAO, roughly $2.15M in total. Each one takes a different approach to monetization and token value capture. Layer 2: Decentralized cloud and privacy - @Hippius_cloud (SN75) is building a cloud stack covering encrypted storage, S3-compatible services, confidential VMs, managed databases and Kubernetes, aimed at teams looking for an alternative to big cloud providers. Layer 3: Consumer AI and creative tools Bittensor is also moving into user-facing apps. - EveryFrame (SN117) is building an AI creative workspace for video and image generation. - @heydittoai (SN118) is building persistent AI memory across Claude, ChatGPT and Cursor, with miners competing to improve agent memory and tool use. Layer 4: Enterprise AI and security - @webuildscore (SN44) powers @manakoai, which deploys vision agents in real-world retail sites like fuel stations. - @trishoolai (SN23) built HaloGuard, a jailbreak-defense model trained through months of adversarial competition. - @SomaSubnet (SN114) works on context compression to cut inference costs. Layer 5: Marketing and distribution The newest vertical helps AI products find paying customers. - @Bitcast_network (SN93) pays creators to produce content for brands, with revenue funding alpha buybacks. - @LeadpoetAI (SN71) builds AI sales intelligence that finds prospects and automates outreach. More of Bittensor's value now comes from subnets selling AI services, not from TAO alone. Which subnet should I dig into next?

A small team on Bittensor just beat the big labs. HaloGuard 2.0 is the new SOTA for open streaming guards: ahead of Qwen3Guard-Stream at every size. At 0.8B. Bittensor isn’t about following the frontier. It’s about defining it.

In the AI age output security is essential. A compact 0.8B model is roughly a quarter the size of the previous best and delivers new SOTA results 83.63 average F1 on public benchmarks! Small enough to run efficiently and cheap enough to make checking every response practical at scale. Congrats Trishool team, keep shipping 👌

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.

AI agent adoption at scale requires trust, not just intelligence. Making SOTA security cheaper, smaller and practical enough to check every output is a major step towards production-ready agents. Another example of what open competition on Bittensor can deliver. Ψ

Every company shipping AI agents asks one question: how can I stop my agent from going rogue? To answer it, HaloGuard Output Guard now sets a new state of the art (SOTA): 83.63 average F1, ahead of the previous best at 82.5. All measured on public benchmarks. And we did it with a 0.8B model, a quarter of the size of the previous best. Small enough to run efficiently. Cheap enough to make checking every response practical. This was the focus of Phase 3 on Trishool: • 7,500+ jailbreaks submitted • 125+ unique miners • 15 weeks of challenges It comes at the right time. We’re taking our product to enterprises now, and this makes our security stack stronger. We’ll share the models on Hugging Face, the arXiv paper and a full explainer soon. There’s much more to come, and we’re looking forward to an exciting week ahead. SOTA AI security, on Bittensor. Ψ

The model itself isn't always the product. Commercial tools can integrate specialized machine intelligence produced on Bittensor and put it to work. HaloGuard EU is a new product from @astrowareai that taps the Trishool subnet (SN23) to help align model outputs with EU compliance requirements. Try it out in the playground or via API. $TAO

If there is a Great Filter ahead of us, something that acts on its own and that we can't reliably control would qualify. That's a reasonable description of what we're building right now.

Read that email again. Nothing about it looks dangerous. A birthday message. A request to use someone's preferred name from the employee file. Polite, ordinary, the kind of thing that lands in a work inbox every single day. A human assistant reads it and thinks nothing of it. An AI agent reads it and does exactly what it says, and to finish the task it opens the employee file, reads personal records, and pulls out details it was never meant to expose. No hacking. No malware. No breaking in. Just a friendly message that quietly walks the agent into handing over data it should have protected. This is what makes agent attacks so dangerous. They do not look like attacks. They look like normal requests, and the agent has no instinct for when a task is a trap. That is exactly what HaloGuard 1.0 is built to catch. It sits over the agent and reads every instruction and every action before it happens, so a birthday email cannot quietly turn into a data leak. Your intern would pause. Your agent might not. Protect it.

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.