Prompt and role manipulation
Test whether malicious instructions, fabricated authority or contaminated context alter the authorised role.
ADVERSARIAL RISK & INTEGRITY ASSESSMENT
ARIA Adversarial Testing
ARIA attempts to discover what the system can be manipulated into doing before that lesson arrives in production.
ASSURANCE LOOP
Unit of analysisSystem + portfolio
DecisionFIT · FIT WITH CONTROLS · NOT FIT YET · SPECIALIST REVIEW
EvidenceVersioned Assurance Case
ActivationHuman authorisation only
Testing tailored to the system, not a universal prompt list.
Tool and permission abuse
Explore improper calls, tool combinations, escalation and composite authority.
Data boundary pressure
Attempt to extract, mix, infer or disclose data outside the authorised purpose and scope.
Memory and context poisoning
Assess how persistent information or earlier instructions can degrade later decisions.
Failure and recovery
Force errors, timeouts, dependency loss and partial states to observe fallback, rollback and containment.
Multi-agent failure paths
Examine chains where several agents produce an action that none controls alone.
EVIDENCE, NOT THEATRE
An adversarial test must be repeatable.
Each finding records preconditions, version, payload, response, effect, severity, linked control and retest result.
Does not proveThat the system is impossible to compromise or safe under every future context.
Does proveWhich version was tested, against which threats, with what results and under which limits.
TriggerChanges to model, prompt, tools, data, permissions or architecture invalidate part of the evidence.
Relationship with Vigilance: ARIA tests the version considered for authorisation before production. Vigilance should later detect changes and operational trajectories that invalidate part of that evidence and require targeted retesting.
A complete cycle, not an isolated audit
Each stage produces a decision, an accountable owner and enough evidence for the next stage not to rely on blind trust.
NeedOrbyn→DesignBlueprint→AssessmentAegis→PressureARIA→AuthorisationHuman→ObservationVigilance→RevalidationPortfolio
NEXT DECISION
Start a conversationThe first question is not what AI can do.
It is which system you are prepared to authorise, under which limits and with what evidence.