Independent assessment of production AI systems: adversarial testing, privacy review, severity-rated findings, and fix verification criteria that produce decision-grade evidence for audit and governance.
AIUC-1 Consortium whitepaper: Defending at Machine Speed, After Mythos (co-author)
Delivered a severity-rated assessment across security, reliability, grounding, privacy, and responsible AI — backed by reproducible test cases, not just a narrative report.
Handed over a prioritized remediation plan that mapped every finding to a control, an owner, and a verification criterion.
The assessment methodology itself became reusable — versioned test packs and evidence formats cut setup time noticeably on every system that came after the first.
Reusable assessment methodology for production AI, adversarial testing, privacy review, severity-rated findings, and evidence formats designed for audit sampling across multiple systems.
Production AI systems carry a combined risk surface across model behavior, retrieval, tool execution, and operational controls. In a regulated environment, stakeholders can't just take design intent on faith — they need an independent assessment of what the system actually does under real usage, under adversarial pressure, and as it changes over time. The output has to be decision-grade: what can ship, under which constraints, with what residual risk, and how a fix actually gets verified. Because these assessments needed to repeat across multiple systems and teams, I built the methodology to be reusable too — versioned test packs, a consistent severity model, evidence formats that hold up for audit sampling without getting rebuilt from scratch each time.
Severity-rated findings register with evidence, reproduction steps, and recommended remediation per finding.
Threat model exists and adversarial test coverage maps to identified abuse cases and trust boundary risks.
Critical and high findings have owners, target dates, and verification criteria; fixes re-tested with closure evidence.
Adversarial testing demonstrates that unauthorized retrieval, exfiltration, and tool misuse are prevented or detected.
Privacy controls validated for runtime behavior, logging retention, and observability access.
Reliability and safety meet defined thresholds; regression coverage established for high-risk change classes.
Assessment dossier complete and sampling-ready with consistent evidence index and control mapping.
| Dataset | Kind | Target |
|---|---|---|
| RAG Grounding & Citation Pack | baseline | Measure faithfulness, grounding quality, and citation coverage. |
| Policy Compliance & Refusal Pack | release regression | Verify refusal behavior for restricted intents and policy constraints. |
| Tool-Use Correctness Pack | baseline | Validate tool selection, argument correctness, and permission boundaries. |
| Adversarial Prompt Injection Pack | red team | Detect susceptibility to prompt injection and jailbreak attempts. |
| Voice Interaction Pack | voice | Evaluate call flows, intent detection, and safety behavior. |
| Sensitive Data Exposure Pack | red team | Detect PII, secrets leakage, and redaction failures across inputs, retrieval, and outputs. |
| Observability and Trace Joinability Pack | audit evidence | Verify traces support investigation and sampling without over collection or broken joins. |
