Autonomous offensive — live targets
ImplementedAutonomous Attack Swarm
A coordinated offensive swarm: N specialized agents attack N live targets in parallel, collaborate by capability, and prove with an observed postcondition.
Family facts
- Code
- FAM-SWARM
- Status
- ready — code maturity
- Availability
- available — whether it is offerable
- Ingest
- live — how a target enters
- Rules of Engagement
- Required (acts on a live target)
- Kind
- Family (a selectable assessment)
What it decides
The same per-finding exploitability truth as RAVEN, proven at the scale of a whole perimeter rather than one target.
How it works
A coordinator orchestrates but runs no security logic: it owns the agent pool, the scheduler, a shared world and the trace emitter, and wires every result to the membrane. The unit of work is the mission instance (a stable id), claimed atomically; an agent whose target is exhausted steals a free in-scope instance from another, so capacity flows to where the work is.
Agents deliberate over one shared world where disagreement is preserved as an explicit contradiction, never resolved by majority vote. Every reported finding crosses the same membrane as a single RAVEN run, so scale never lowers the bar for proof.
A worked example
From one order over several authorized live targets, the runner seeds one RAVEN drone per target with active Rules of Engagement; findings from all drones converge on the membrane and only verified ones are counted.
Declared limits
Runs coordinated RAVEN agents over live targets, verified end-to-end, in a single process; multi-target fan-out from a single order needs multi-input ingest wiring. Diverse per-agent priors are not wired yet — today it splits territory by target, not by hypothesis.
References
- [1]Blackboard systems (Hearsay-II)
- [2]Contract-net / work-stealing schedulers
References point to public, well-known technical sources for further study.
Related articles