Subnet345 published a controlled evaluation conducted between the 25th and 28th of July 2026 by the UK AI Security Institute (AISI) testing frontier models including Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol with cyber classifiers disabled.

Subnet345 is a software company for the era of agents. We build the accountability layer for AI agents: an operating record of what agents did, running inside an organization's own perimeter, on its own terms. As teams hand real work to fleets of agents, the question that matters stops being what a model can do and becomes what its agents actually did, on whose authority, and whether you can prove it after the fact. Most of the field is racing to build the capability. We build the record beneath it. We build from a single conviction: in the era of agents, the bottleneck is not intelligence, it is accountability, and accountability has to be built as its own layer rather than reconstructed after something goes wrong. We run the way we build. Subnet345 operates its own work as a multi-agent swarm on the substrate we build, with human accountability at the practice level, so the software is proven against our own operations first. We contribute to the open-source foundations we rely on, including the message bus at the core of the substrate. We build for the environments where the answer has to hold up to an examiner: regulated, security-sensitive, and sovereignty-minded operations. The software installs where you run, and you verify the record yourself, inside your boundary. Three properties run through everything we build: memory, accountability, coordination. Est. 2026.

Founded 2026