PHAGE

Inspiration

Every guide to securing AI agents describes the same shape: put a guardrail in front of the model and block known-bad input. That works until an attacker rewords the attack. It also tells you nothing about which of your agents are actually vulnerable, or to what.

Biology solved this differently. An immune system doesn't只 filter — it gets infected, learns the pathogen, and recognises it on return. We wanted to know whether a fleet of agents could work the same way: attack itself on purpose, measure what actually lands, and build a memory of it.

Built With

• agent-development-kit
• agent-engine
• cloud-run
• fastapi
• gemini
• gemma
• google-cloud
• memory-bank
• opentelemetry
• python
• sqlite
• vertex-ai