Full project description
Phylax is a decentralised trust layer for the AI software supply chain. It analyses the third-party artifacts and dependencies that software systems and autonomous agents rely on (agent skills, MCP servers, packages, and repositories) and returns a signed, verifiable safety verdict on each, blocking anything dangerous before it reaches the environment.
Live now: https://app.phyi.dev
HOW DETECTION WORKS
Detection is behavioural, not signature-based. For each artifact, Phylax derives what it claims to do (its declared intent), runs it in an isolated, instrumented sandbox, and traces its behaviour step by step. Any deviation between declared intent and observed behaviour is the signal of malice. This catches novel, hidden, and obfuscated threats that known-flaw databases miss, including multi-step attacks that look harmless in isolation.
HOW THE NETWORK WORKS
Detection quality is produced by open competition on a dedicated Bittensor subnet. Each round:
Miners build the detectors. Miners build security agents and submit them to the network as hash-pinned artifacts, so the exact agent evaluated is provably the one submitted. They improve and resubmit between rounds.
The chain selects the test set. Artifacts are drawn deterministically from a large labelled benchmark of known-safe and known-malicious samples, seeded from the chain, so every validator derives the same set with no coordinator and no cherry-picking. A held-out portion is never scored, preventing memorisation.
Validators score against ground truth. Validators run each agent against the selected artifacts in instrumented sandboxes and score it on how accurately its verdicts match the known answers.
Consensus sets the ranking. Validators assign on-chain weights from these scores, and Yuma Consensus combines them via a stake-weighted median, so the ranking reflects the staked majority, not any single validator, and dishonest scoring is penalised.
The best agent wins. Emissions reward the highest-scoring miners, and the top agent per artifact type becomes the champion that serves live verdicts.
VERIFIABLE BY DESIGN
Every verdict is written on-chain as a signed attestation anyone can verify. This is the decisive difference from centralised tools: they ask you to trust their team, Phylax lets you verify the result yourself
Why it works on Bittensor
A security verdict you have to trust is weak; a security verdict you can verify is strong. That's only possible if no single company owns the judgment, which is exactly what Bittensor gives us. Detection quality is produced by a decentralised network of miners competing to build the most accurate security agents, rewarded only when they win. Quality improves continuously through competition rather than at the pace of one company's payroll, and every verdict is signed and written on-chain so anyone can verify it. Phylax isn't a Web2 product bolted onto a token, it needs decentralisation to make its core promise real