AI red teaming
Aplark Red
Find the weaknesses before the field does.
Aplark Red is intended to search for the inputs and conditions under which a system fails, and to record those failures as structured evidence against the hazard register rather than as a list of interesting images.
The problem
Adversaries do not sample uniformly
Robustness measured against random perturbation says little about robustness against a party choosing the perturbation. Mission-critical systems operate against adversaries who will find the worst case, and a test regime that samples the operational domain uniformly will not.
Red teaming produces findings. What certification needs is findings bound to hazards, with the search procedure itself recorded well enough that someone can argue about whether it was thorough.
How it works
How it is intended to work
Aplark Red is being designed to run adversarial search as a campaign against a declared threat model, emitting each discovered failure as evidence bound to the requirement it defeats, at a recorded reproducibility class.
The search procedure, budget and termination condition would themselves be recorded, because "we red teamed it" is not a claim a reviewer can check and "we searched this space under this budget and found these failures" is.
Capabilities
What it does
Threat-model-scoped search
Planned to take a declared threat model and perturbation budget as the campaign definition, so that findings are interpretable against a stated adversary.
Findings bound to hazards
Intended to emit each failure against the requirement and hazard it defeats, in the evidence format the assurance case consumes.
Recorded search procedure
Designed so that the search itself is an auditable artefact — budget, method and termination condition — not an unrepeatable exercise.
Who it is for
- Programmes whose systems will operate against a capable adversary
- Safety engineers who need robustness claims that survive challenge