Palisade

The evaluation rig measures what matters for an agent, not a chatbot: how often an attack hijacks a tool call and how often it alters the model's stance, against a fake-tool server with session-isolated sandbox state, so nothing adversarial ever has a real side effect. The arena runs red team against blue team in self-play, with a self-hardening persona that rewrites its own prompt until attack success collapses.

The first deployed consumer is a live defense inside coding-agent sessions: hooks screen incoming content, flag offenses per session and close the gate until a person reviews them. Every input is scored deterministically; a model may raise suspicion but never clear it. The reusable library, the HTTP gateway and the MCP packaging are designed and not yet built.