AI security and risk assessments
Find the AI already running in your organization, work out what it can reach, and test whether its guardrails survive contact with an adversarial user. Governance mapped to the NIST AI Risk Management Framework, testing mapped to OWASP and MITRE ATLAS.
Bought under NAICS 541511 and 541512.
What this service does
Two different questions get bundled together as "AI risk", and they need different work. The governance question is whether you can account for the AI in your estate, what data it touches, and who approved it. The security question is whether a deployed AI system can be made to do something it should not when someone actively tries.
We answer both. Governance findings map to the Govern, Map, Measure and Manage functions of the NIST AI Risk Management Framework, with the Generative AI Profile applied where relevant. Security findings come from testing a live system against the OWASP Top 10 for LLM Applications categories and MITRE ATLAS techniques.
The inventory usually comes first, and it is usually the uncomfortable part.
What we assess and protect
- Sanctioned AI platforms and the data they are connected to
- Shadow AI: tools adopted without approval, and vendor features switched on by default
- Retrieval sources and whether a model can reach regulated data
- Model hosting: third-party API, private hosting, or air-gapped deployment
- Agents and tool-calling: what the system is permitted to invoke
- Guardrails, output handling, and authorization checks
- Prompt and response logging, retention, and audit readiness
How the work runs
Select any stage in the diagram to see what happens there. The same model drives the engagement plan you receive.
Select any cell for what the weakness is, how we test it, and what it costs you. Play a scenario to watch one exploit move through the components. Levels are illustrative, not findings.
Inventory and discovery
Sanctioned platforms, embedded vendor features, and what staff adopted independently. We use network, identity, and procurement signals, because asking people what they use does not find shadow AI.
Data-flow mapping
What each system can retrieve, what it was trained or fine-tuned on, and where prompts and responses are stored. Where regulated data enters a model, the obligation follows it.
Governance assessment
Scored against the NIST AI Risk Management Framework functions, with the Generative AI Profile applied to generative systems.
Adversarial testing
Direct and indirect prompt injection, data leakage, insecure output handling, excessive agency, and tool abuse against the deployed system, not against a description of it.
Guardrail verification
We test whether the filter, the authorization check, and the rate limit hold under adversarial input, rather than reviewing the policy that says they exist.
Monitoring and policy
Logging design, retention aligned to your obligations, alerting on the abuse patterns we found, and an acceptable-use policy people can actually follow.
What you receive
- AI inventory and risk register covering sanctioned and shadow systems
- Data-flow map showing what each system can reach
- NIST AI RMF-aligned governance assessment with scoring
- Adversarial test findings with reproduction steps
- Guardrail verification results
- Logging, retention, and monitoring design
- AI acceptable-use policy and approval workflow
If the answer is to keep the data in-house
For some organizations the cleanest resolution to AI data risk is not a policy, it is architecture: run the model inside your own boundary so the data never leaves it.
Swif7Edge builds that as a product. Swif7AI is a private model stack in three sizes with fixed-cost deployment, audit-ready logging of every prompt and response, and model selection you control rather than a vendor controlling it for you.
Frameworks we work to
Named, published, and checkable. Where a project revises frequently we cite the framework and confirm the current revision with you at kickoff rather than printing a version number that will age badly.
NIST AI RMF 1.0
Published January 2023. The Govern, Map, Measure and Manage structure that organizes every governance finding we report.
NIST AI 600-1
The Generative AI Profile, published July 2024. Applied wherever the system in scope is generative.
OWASP & MITRE ATLAS
The OWASP Top 10 for LLM Applications categories and MITRE ATLAS adversarial techniques drive the testing plan.
Who performs the work
Swif7Edge is small on purpose. The people who scope your engagement are the people who run it.
Our practice was founded by U.S. Army veterans. The security side is led by a combat veteran and former NSA contractor with 26 years in security engineering, who has built and led teams of more than 130 engineers, stood up global security operations centers, and directed offensive security work for the Defense Intelligence Agency, the FBI, and the U.S. Marine Corps.
That background is company-stated. Specific engagements are covered by non-disclosure terms and in most cases by the sensitivity of the work itself, so we do not detail them publicly. We will substantiate it directly during a procurement, with references and whatever contract detail can be released.
The strategy side brings more than 20 years across the Intelligence Community, the Department of Defense, and Fortune 500 operations, managing portfolios above $150 million and cross-functional teams of up to 2,500 personnel.
Credentials held on the team
- Offensive Security Certified Professional (OSCP)
- Certified Information Systems Security Professional (CISSP)
- GIAC Penetration Tester (GPEN)
- GIAC Certified Incident Handler (GCIH)
- GIAC Assessing and Auditing Wireless Networks (GAWN)
AI assurance work draws on the same offensive discipline. Prompt injection and tool abuse are exploitation problems before they are governance problems, which is why the testing is run by the same operators.
We do not publish client names, logos, or case studies. The work is mostly covered by non-disclosure terms, and a small firm that lists logos it cannot substantiate is making a claim it cannot stand behind. The certifications and registrations above are independently checkable, and the methodologies on each service page are named rather than described. Everything we say about our own history is company-stated and we label it that way.
Scope it before you budget it
Tell us the environment and the obligation. We will come back with the approach, the duration, and what drives the price — before anyone asks you for a number.
Common questions
If what you need is not here, ask us directly rather than filling in a form.
There are named frameworks and we work to them. The NIST AI Risk Management Framework (AI RMF 1.0, published January 2023) provides the Govern, Map, Measure and Manage structure, and NIST AI 600-1, the Generative AI Profile published in July 2024, extends it to generative systems. For adversarial testing we work to the OWASP Top 10 for LLM Applications categories and MITRE ATLAS. We cite the framework rather than a version number for the latter two, because those projects revise frequently and we would rather be accurate than precise.
Vendor self-attestation is not third-party verification, and it rarely covers how the product behaves once connected to your data and your tools. The risk usually lives in the integration: what the model can retrieve, what tools it can invoke, and what happens when a user or a document instructs it to do something it should refuse.
Shadow AI is the AI tooling staff adopted without telling anyone, plus the AI features vendors switched on inside products you already licensed. In most assessments the inventory is itself the main finding, because you cannot govern or secure a system nobody has recorded.