Maity: Board

Idea (later): a decision model as a guard that filters dangerous requests

tsk-3m5spzcrnd0y proposed

Oliver, 2026-10-29 (for later): a decision model acts as a guard and filters dangerous requests.
Coordinator's reading, to confirm when it comes up: before a request (§7: decision, consent, an action a proxy asks the owner to approve) or an agent-to-agent/agent-to-human message reaches its addressee, a guard model scores it (e.g. destructive operations, secret exfiltration, prompt injection, social engineering of a human); a dangerous one is held and shown to the owner with the reason instead of delivered, a doubtful one is delivered with a warning. Related idea: tsk-5xqzz2scmebr (guard checks messages between agents, premium). Open: which model, where it runs (hub, operator, external API), cost and privacy (what leaves the instance), false positives, and that it never replaces the hub's own checks.