Trishool is an innovative AI alignment protocol built on the Bittensor network, designed to ensure that advanced AI systems remain aligned with human values. By leveraging a decentralized network of miners and validators, Trishool provides a robust mechanism for adversarial training, allowing for the continuous improvement of AI guard models. This protocol addresses the critical need for security in AI by enabling real-time responses to emerging threats, making superintelligence safe for humanity.
Miners in the Trishool subnet submit adversarial prompts targeting AI agents, which are then evaluated by validators using specialized scoring criteria. Successful exploits retrain the guard models, ensuring they adapt to new challenges every week. This collaborative effort creates a self-evolving network that hardens AI security in real-time.
Trishool aims to be the leading solution for AI safety, with a long-term goal of creating a universally trusted framework for AI alignment. By fostering a global community of researchers and engineers, Trishool seeks to significantly reduce the risks associated with advanced AI systems and promote their safe deployment across various applications.
Team information not publicly disclosed.
Trishool runs continuous adversarial competitions where miners attempt to jailbreak AI models. Every successful attack trains the model to be stronger. After 14 weeks of non-stop competitive jailbreaking, their 0.8B parameter model achieves ~90% resistance to known attacks, targeting 99-100%. The approach mirrors Anthropic's Constitutional AI — define what's safe, train against violations, harden continuously. Featured on the OTF Novelty Search (Jun 26, 2026) where Const confirmed he's personally using their GM tooling.
Agent security is a massive market — every company deploying AI agents needs input/output guards. Trishool is first-to-market on Bittensor with Constitutional AI + adversarial training. Const personally endorsed them on OTF call and is using their tooling. The 0.8B model runs on phones — zero infrastructure cost for deployment. If they hit 100% jailbreak resistance, this becomes the industry standard.
Low active miner count means the adversarial competition pool is small. The model is only 0.8B parameters — larger models may require different approaches. Revenue model depends on enterprise sales which are slow cycles. Competition from Anthropic's own Constitutional AI classifiers and major cloud providers building safety layers.
Three moats: (1) Decentralized adversarial training — thousands of attackers continuously probing the model, generating attack data no single company could produce. (2) Constitutional AI approach that lets customers define their own safety policies on top. (3) Tiny model size (0.8B) means near-zero latency as a guardrail layer — can sit in front of any LLM without performance impact. No centralized competitor can match the breadth of adversarial data from a competitive jailbreaking subnet.
AI agent security is exploding as agents get access to databases, wallets, and production systems. Trishool is the only subnet building dedicated guardrails for this. Enterprise clients are already using the product. With price-based emissions rewarding real builders, Trishool's +53% 30-day price momentum reflects genuine demand, not hype
"The only subnet on Bittensor where miners literally try to break AI safety and the model gets stronger from every attack. Const is using their tooling. Enterprise clients are buying. This is real."