AI Security & Architecture Review
I review how an AI system crosses trust boundaries: which data enters the model, which tools it can call, what those tools can change and how the result is verified before it affects production or a customer.
What is reviewed
The review maps the agent, models, prompts, retrieval sources, credentials, tool permissions and external services that participate in the workflow. It looks for places where untrusted content can influence instructions, where permissions are broader than the task requires, and where a failure can pass through without a reliable signal.
How the scope is defined
We start with the current architecture, a representative workflow and the outcome you need. The first stage is bounded around agreed components and environments. The goal is to produce findings that can be reproduced and acted on, not a generic checklist detached from the implementation.
What you receive
The delivery can include a trust-boundary and data-flow description, prioritized findings, concrete remediation options and verification steps. Where implementation is included, changes are checked against agreed acceptance criteria and handed over with the assumptions and remaining risks documented.
What the review does not claim
A scoped review does not prove that a system is free of vulnerabilities. Coverage depends on the supplied code, configuration, access and test environment. Findings distinguish observed evidence from inference, and residual risk is stated explicitly. Greater autonomy should follow narrower permissions and stronger verification, not precede them.