AI Transparency Testing

Identical candidate profiles with demographic variables (name, gender indicators) swapped to detect differential scoring or recommendation patterns.

Methodology: 400 pairwise flip tests across 8 job types. 800 API calls against GPT-5.2.

Structured adversarial prompts designed to surface differential responses based on user demographic signals embedded in conversation context.

Methodology: PAIR (Prompt Automatic Iterative Refinement) methodology. 500+ prompt variants per agent.

Independent third-party penetration testing covering prompt injection, jailbreak attempts, data exfiltration, and attempts to extract system instructions across all AI agents.

Methodology: Annual third-party penetration test conducted by an accredited security provider. Summary findings and remediation status available via our Trust Center.

Evaluation of the agents' ability to identify expressions of risk of harm to self, including passive ideation, active ideation, and crisis signals, and respond with appropriate safety messaging and escalation.

Methodology: Automated script testing 100 graded-severity vignettes. Each response scored against a predefined rubric for detection and response quality. Pass rates reported per severity tier.

Assessment of whether agents correctly identify expressions that may indicate delusional ideation or psychotic features, and respond with appropriate safety messaging rather than reinforcing or elaborating.

Methodology: 100 vignettes per agent across 5 tiers. Psychosis-specific rubric (Recognition, Non-Reinforcement, Empathy, Professional Referral, No-Harm) with 4-strategy judge fallback.

Testing whether agents identify contextual safeguarding risk signals (e.g. disclosure of alcohol use while responsible for the care of children) and provide appropriate guidance without overstepping scope.

Methodology: 80 vignettes across 4 domains. Safeguarding rubric (Recognition, Sensitivity, Appropriate Guidance, Scope Awareness, No-Harm) with 4-strategy judge fallback.