A New Milestone for AI Safety as Trishool (SN23)’s Output Guard Nears SOTA

Trishool (SN23)’s new AI safety system is closing in on leading performance after making a significant jump in just two weeks.

SN23’s Output Guard has risen from a 73.9% F1 score at launch to 78.53%, closing roughly half the remaining gap to the current state of the art. The improvement comes as the system faces continuous testing from participants across the Bittensor network.

AI safety does not stop when a user sends a prompt. A model can still produce harmful content, expose sensitive information, or attempt a dangerous action in its response, which is where an output guard comes in.

An output guard monitors what an AI model produces and is designed to catch problems as they emerge.

Trishool is building its Output Guard to detect harmful responses, risky actions, and potential data leaks before they reach users or are carried out.

As a project focused on AI alignment and safety, Trishool (SN23)’s latest progress shows how quickly the system is developing under continuous testing from the network.

The initial 73.9% result was described as a starting point, and the latest benchmark represents a 4.63-point improvement. It has also brought the system roughly halfway toward closing its remaining gap with leading models.

That progress is being driven through ongoing challenges on Bittensor, where participants continually test the guard for weaknesses. Successful attacks can expose new problems and create fresh examples that help improve the system.

The approach follows Trishool’s Input Guard, which the subnet says has already reached the top of its category (SOTA). The Output Guard is now following a similar upward trajectory as testing continues.

Trishool also recently announced that it had joined OpenAI’s cyber program around the time the Output Guard launched. For now, the clearest sign of progress is how quickly the system has advanced since its initial release.