AI Infrastructure Audit.

An AI infrastructure security audit from XSource Security is a source-level review of the code your AI product runs on, backed by live testing of the running build. It covers the serving stack, the retrieval and agent layer and the accelerator runtime underneath, and every finding is reproduced before it is written up.

What is in scope

• Request parsing, multi-tenant cache and session isolation, resource exhaustion in inference servers and gateways
• Prompt and retrieval paths: cross-user exposure, injection, output handling
• Agents and MCP: tool permissions, memory, multi-step attack chains
• Accelerator runtimes and the drivers between the workload and the hardware
• Penetration testing of the web and API layer around the model, with authenticated roles

Who it is for

Teams shipping an AI product on top of an open-source serving stack, an LLM gateway or an application platform, and teams building agents that hold real credentials and tools. The review covers your own code and the way the upstream components are wired together, which is where most of the published findings in this area were.

How the engagement runs

1. Scope and threat model. The component, its trust boundaries, the attacker positions that matter and the exact build under review are fixed in writing before any code is read.
2. Map the surface. Entry points, parsers, privilege transitions and the past security fixes in that code, because incomplete fixes are where the variants live.
3. Reproduce. Tooling produces candidates; a candidate becomes a finding only when it runs, from the attacker position the scope names.
4. Report, disclose, retest. Advisory-grade write-ups with evidence, CVSS rationale and remediation. One retest within 60 days is included.

One researcher runs each engagement end to end: scoping, review, reproducing every finding and writing the report. Scope, access, findings, questions and retest results are handled in writing.

What you receive

• Confirmed scope, threat model and methodology, agreed in writing before any code is read
• Prioritized findings, each reproduced live against the build in scope, with a working proof of concept
• CVSS with rationale and the attacker position each finding requires
• Written remediation naming the location and the change
• An executive summary written for a reader who will not open the technical section
• One asynchronous retest within 60 days
• Coordinated disclosure and the CVE request with the vendor or CNA, on request, where a finding belongs upstream

The report is written to be handed straight to whoever asked for it: customer security questionnaires, SOC 2 / ISO 27001 evidence requests, ISO/IEC 42001 and NIST AI RMF work for AI systems, procurement or investor review.