Welcome to Phylax

Phylax is a Bittensor subnet that turns untrusted AI agent skill bundles into Signed Skill Safety Attestations. Validators dispatch skill bundles to miners. Miners run their analysis pipelines and return signed attestations. Validators verify those submissions, compare them across a verification group, and push the resulting weights on-chain. The goal of this documentation is to explain every part of the subnet in enough detail that you can run a miner, run a validator, or understand how the network rewards honest analysis.

​

Setup Guides

​

Learn the Subnet

​

What is Phylax for

Agent ecosystems scale through composable skills. The same composability creates a large attack surface. A malicious skill can steal API keys, exfiltrate data, establish persistence, or hijack agent behaviour through prompt injection. Existing scanners use an LLM as the analyser, which is fast to build but fundamentally limited. The model itself is prompt injectable. Evasion through obfuscation is trivial without real execution. The output is plain text rather than an enforceable contract. Phylax runs a decentralized competition where miners perform real behavioural sandbox detonation, produce cryptographically signed evidence packs, and emit machine readable policies that runtimes can enforce. Validators independently verify those submissions and compute consensus across a verification group, so a single miner cannot fake a result and a group of colluding miners cannot fabricate the per task probe events the bundle needs to actually execute.

​

How to read these docs

If you want to run something, jump directly into the Miner Guide or Validator Guide. Both are step by step setup paths. If you want to understand how the network works, start with the Subnet Overview, then read Incentive Mechanism and Verification Groups. If you want a reference on a specific schema or wire format, jump into Core Components or Reference.