# Multi-agent threat model (Evolution 15)

In scope: compromised or revoked delegates, forged ancestry, escalation through chains, impostor agents, replayed and
forged agent requests, colluding agents amplifying one source, self-approval, poisoned shared memory, stale plans
after revocation, and policy change mid-run.

Residual risk: software cannot prove that two agent keys, two observers or two approvers are controlled by different
people. Independence of evidence roots is as good as the operator's assignment of observers to roots. Agents that
share one upstream source and present it through two genuinely separate observers will be counted twice. Anything not
routed through the runtime is outside it.
