Trishool’s Bittensor AI Safety Subnet Joins Google for Startups Web3 Program

Trishool, the decentralized AI alignment subnet operating as SN23 on Bittensor, has been accepted into the Google for Startups Web3 Program. The participation includes up to $200,000 in Google Cloud credits over two years, $12,000 in Enhanced Support credits, twelve months of Google Workspace Business Plus, access to a dedicated Startup Success Manager, and entry into Google's gated Web3 engineering community.

Trishool runs adversarial evaluations against frontier large language models at scale, making infrastructure runway its core constraint. Miners in the subnet submit seed prompts designed to probe models for deception, sycophancy, manipulation, and power-seeking behavior, with outputs scored through a Petri-based evaluation framework. Additional GPU-hours directly expand the network’s coverage of the AI safety manifold it is mapping.

The team positions this as an answer to the rising cost and frequency demands of model audits, which can run into six figures per evaluation cycle while new frontier models continue to ship. They highlight that recent regulations, including fines up to €35 million for non-compliance under the EU AI Act and independent audit mandates for advanced systems under California’s SB-53, are tightening oversight expectations for high-capability models.

Program benefits go beyond credits, with early access to Google Cloud’s Web3 product roadmap, a gated Discord with Google’s Web3 engineering teams, hands-on technical labs, and exclusive grants from ecosystem partners. Trishool frames this as key distribution and integration surface for building safety infrastructure that other Bittensor subnets and external enterprises can consume downstream.

The announcement follows Trishool’s integration with Chutes (SN64), which onboarded Halo Guard Alpha, Trishool’s safety classifier, into the Chutes inference stack and Chutes Chat. The team presents this as an example of Bittensor’s intended pattern of cross-subnet composability, with alignment running on top of decentralized compute. They summarize their positioning by stating that safe superintelligence is the core objective, and that the partnership effectively extends their runway to pursue that goal with support from a major cloud provider.