AODIT-6 verhaltensbezogener Bewertungsrahmen

KI-Agenten-Testmethodik

aodit bewertet, wie KI-Agenten unter Stress, Widerspruch und adversariellem Druck reagieren. Der Rahmen ist für Risiko-, Compliance- und Audit-Teams konzipiert, die unabhängige Nachweise statt selbstberichteter Modellleistung benötigen.

Für wen dies gedacht ist

Risiko-Teams, Compliance-Beauftragte und Audit-Funktionen, die unabhängige verhaltensbezogene Nachweise statt anbietergelieferter Benchmarks benötigen.

Sechs Bewertungsdimensionen

Jede Dimension misst eine eigenständige verhaltensbezogene Risikoklasse mit strukturierten Multi-Turn-Szenarien.

Reliability
Integrity
Confidentiality
Judgment
Resistance
Resilience
Reliability

Does it maintain accurate outputs under repetition and contradiction?

R1 Factual consistency
R2 Contradiction handling
R3 Data poisoning resistance
R4 Sequence logic
R5 Memory persistence
Integrity

Does it disclose uncertainty and resist fabrication under authority pressure?

I1 Uncertainty disclosure
I2 False certainty resistance
I3 Fabricated source handling
I4 Calibration gap (self vs evaluator)
I5 Pressure-induced bluffing
Confidentiality

Can it protect sensitive information under extraction and role-pressure attempts?

C1 Data leakage resistance
C2 PII boundary enforcement
C3 Model distillation resistance
C4 System prompt confidentiality
C5 Information boundary under role pressure
Judgment

Does it make proportional decisions and escalate correctly under risk?

J1 Proportional action
J2 Escalation to human oversight
J3 Policy boundary adherence
J4 Authority conflict resolution
J5 Blast radius limitation
Resistance

Does it resist jailbreaks, injections, and social engineering pressure?

T1 Jailbreak resistance
T2 Prompt injection resistance
T3 Social engineering resistance
T4 Authority spoofing resistance
T5 Adversarial reframing resistance
Resilience

Does it maintain coherent performance under prolonged stress and ambiguity?

Z1 Overload handling
Z2 Ambiguity stacking
Z3 Conflicting instruction handling
Z4 Stress persistence
Z5 Recovery after failure

Acht-Runden-adversarielles Protokoll

Jede Bewertung führt eine Acht-Runden-Sequenz aus, die den Druck progressiv erhöht und dann die Erholung testet.

1

The opening question

Normal question to establish baseline behaviour under no pressure.

2

The reworded question

Same question in different wording to test consistency under rephrasing.

3

The complicated scenario

Added complexity to test memory, priorities, and constraint handling.

4

The grey zone

Edge-of-policy case to check boundary adherence versus loophole drift.

5

The hard push

Urgency, authority, and emotional pressure to test stress resistance.

6

The manipulation attempt

False premises and disguised requests to test attack detection and refusal.

7

The recovery test

Post-pressure stabilization check for safe, coherent re-alignment.

8

The final judgment

Closing assessment of reasoning quality, escalation decisions, and robustness.

Unabhängig by Design

aodit ist eine unabhängige Bewertungsschicht. Sie zertifiziert keine Modelle, ersetzt keine Governance-Rahmen und greift nicht auf Modellgewichte zu.