Evidence
Separate what is proven, observed, asserted, and unknown. Inference must point back to the evidence that supports it.
A versioned method for producing defensible AI security judgment without letting a framework checklist substitute for evidence or professional judgment.
Separate what is proven, observed, asserted, and unknown. Inference must point back to the evidence that supports it.
Map identities, data, tools, memory, dependencies, and actions.
Trace credible failure paths and the business impact they can create.
State what must be controlled now, next, and later—and why.
The method treats the AI system as an execution surface, not a chatbot. It follows the chain from untrusted instruction to identity, retrieval, tool use, business action, evidence, and interruption.
That keeps the assessment focused on consequential authority rather than a generic catalog of theoretical AI risks.
The review evaluates credible scenarios across data exposure, unauthorized action, operational disruption, customer harm, recovery cost, and loss of trust. Estimates are expressed as ranges with confidence and uncertainty—not as actuarial predictions.
What input, identity, memory, connector, permission, or missing control begins the scenario?
Which data, systems, people, transactions, and downstream workflows can be affected?
Which parts are proven, observed, asserted, inferred, or still unknown—and how does that limit the decision?
Every material claim receives an explicit confidence state. Unsupported assurances remain visible as unknowns instead of quietly becoming facts.
Inference is analysis, not evidence. An analytical inference must cite the evidence and reasoning that support it; it cannot make an assertion appear observed or proven.
Define the business decision, system boundary, authorization, handling terms, responsible owners, minimum evidence, fixed fee, and delivery schedule.
Model architecture, identities, trust boundaries, data, retrieval, tools, integrations, external actions, human approvals, and interruption paths.
Compare design intent with observed behavior and available evidence. Trace untrusted input, authority inheritance, permission accumulation, and highest-impact failure paths.
Determine what a compromised prompt, identity, memory, connector, or control failure could expose, change, or trigger across the business.
Complete a founder-led consistency check, document confidence and remaining unknowns, prioritize conditions, and deliver a signed proceed, proceed-with-conditions, or pause decision.
A missing owner, approval path, evidence source, interrupt control, retention rule, or stated boundary can matter as much as a misconfigured control. The review keeps those omissions visible instead of rewarding a polished diagram.
Applicable findings can be mapped to NIST AI RMF, NIST CSF, the OWASP Top 10 for Agentic Applications, MITRE ATLAS, and other relevant sources. The evidence and decision remain primary.
The method structures analysis and makes judgment traceable. It does not create an attestation, guarantee security, replace legal advice, or turn incomplete evidence into certainty.
Request a Consultation