Skip to main content
AI // Security & governance

AI your security team can sign off on.

From private model deployments to prompt-injection defenses and audit logs, we design AI systems with the same discipline as any system that touches sensitive data.

inference.tracelive

{

"summary"
: "Know what data goes where, who can ask what, and keep a record of every answer.",
"hosting"
: "Private or approved vendors",
"controls"
: "Access · redaction · logging",
"threats"
: "Injection · leakage · misuse",
"reviewed_by"
: "a human"

}

Common AI risks and controls

RiskExampleControl
Data leakageSensitive data sent to an external APIRedaction, private models, vendor terms review
Prompt injectionA document tells the model to ignore its rulesTreat content as data, restrict tools, validate outputs
Over-permissioned toolsAgent can delete recordsScoped credentials, approval gates
Untraceable answersNobody knows why it said thatFull prompt and response logging
Model driftVendor updates change behaviorPinned versions, regression evaluations

Our governance baseline

  • Data inventory: what each AI feature can see
  • Approved model and vendor list
  • Role-based access to AI features
  • Logging with retention aligned to policy
  • Evaluation and red-team tests before release
  • A usage policy staff actually understand

Why AI needs its own security thinking

AI systems introduce risks that traditional application security doesn't fully cover. Models can be manipulated by instructions hidden in documents or emails. Prompts can carry sensitive data to an external service. Agents with tool access can be talked into actions their designers never intended. And because model outputs vary, a system that passed its tests yesterday can behave differently after a vendor update.

None of this means avoiding AI. It means treating AI features like any other system that touches sensitive data: knowing what data flows where, limiting what each component can do, logging what happened, and testing for the specific ways AI systems fail.

We help teams put those controls in place without slowing every project down, so the security review becomes a checklist rather than a blocker.

Controls we put in place

Data minimization

Each AI feature sees only the fields its job needs; sensitive values are masked or redacted.

Private deployment options

Open-weight models hosted on your infrastructure when data must not leave it.

Injection defenses

External content treated as data, tool use restricted by context, and outputs validated.

Access control

AI features respect existing roles, including in retrieval and query results.

Audit logging

Prompts, retrieved sources, tool calls, and responses logged with retention rules.

Evaluation and red-teaming

Adversarial tests before release and after model or prompt changes.

A governance engagement

  1. Step 01

    Inventory

    Week 1

    List the AI tools and features in use, official and unofficial, and the data they touch.

  2. Step 02

    Risk review

    Week 1–2

    Assess each against data sensitivity, exposure, and potential harm.

  3. Step 03

    Controls

    Week 2–3

    Design technical controls and approval rules proportionate to the risk.

  4. Step 04

    Policy

    Week 3

    Write a practical usage policy and an approved-tools list with your legal team.

  5. Step 05

    Testing

    Week 3–4

    Run adversarial tests on the highest-risk systems and fix what they find.

Hosting options compared

OptionData leaves your network?Trade-off
Hosted API, standard termsYesFastest start; check retention and training terms
Hosted API, enterprise termsYes, under contractStronger guarantees; still external processing
Cloud model in your tenancyStays in your cloud accountMore setup; good balance of control and quality
Self-hosted open-weight modelNoFull control; you run and patch the infrastructure

Questions your security team will ask

  • Which data is sent to which model, and where is it processed?
  • Is anything retained by the vendor, and for how long?
  • Who can use each AI feature, and what can it access?
  • How are prompt injection and data exfiltration tested?
  • What is logged, and who can read the logs?
  • How are model and prompt changes reviewed?

Common pitfalls we help you avoid

The first pitfall is a policy without tools. Telling staff not to paste customer data into public chatbots works poorly unless an approved alternative exists that is just as easy to use.

The second is reviewing each AI project from scratch. A standard set of controls and a short risk checklist let low-risk projects move quickly while high-risk ones get deeper review.

The third is assuming a vendor's defaults are safe. Retention, training on customer data, and processing regions vary by product and plan, so we check terms against your requirements before anything goes live.

What we measure

MeasureWhy it matters
AI systems inventoriedCoverage of tools in use, including unofficial ones
Controls in placeLogging, access, and data rules per system
Test findingsIssues found in adversarial testing and their status
Policy adoptionStaff awareness and use of approved tools

Who does what

We handle

  • Inventory, risk review, and control design
  • Technical controls and testing
  • Policy drafting support

Your team handles

  • Final policy approval and legal review
  • Owners for each AI system
  • Ongoing enforcement and training

A worked example

Consider a company where staff already use several AI tools, some approved and some not. An inventory finds customer data being pasted into a public chatbot and an internal assistant with access to every document. Approved tools are chosen, the assistant's retrieval is limited by role, and logging is added.

A short policy explains what is allowed and why, and the security team receives a standard checklist for future AI projects, so the next one is reviewed in days instead of weeks.

What success looks like

Success is AI adoption that speeds up rather than stalls: approved tools people actually use, clear rules, and security reviews that take days instead of weeks.

Frameworks we map controls to

OWASP's Top 10 for LLM Applications 2025 covers prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation, and unbounded consumptionSource [1]. It's a practical checklist for threat modelling an AI feature.

For governance, NIST's AI Risk Management Framework, released in January 2023 for voluntary use, and its Generative AI Profile from July 2024 give a structure your risk and compliance teams will recognizeSource [2]. Controls such as minimal permissions and human approval for high-impact actions come straight from OWASP's guidance on excessive agencySource [3].

Frequently asked questions

Can we use AI without sending data outside our network?
Yes, with self-hosted open-weight models on your infrastructure.
Do hosted AI vendors train on our data?
It depends on the vendor and plan. We review terms with you and choose options that match your policy.
How do you test for prompt injection?
With adversarial test cases that try to override instructions, exfiltrate data, or misuse tools.
Can you help write an AI usage policy?
Yes, as a practical document tied to the tools you actually use, reviewed by your legal team.
Is this only for large companies?
No. Small teams handling customer or financial data need the same basics.
Do you map controls to frameworks?
Yes. We can align controls with frameworks your organization already uses, such as ISO 27001 or the NIST AI Risk Management Framework, alongside your auditors.
What is shadow AI and should we worry?
It is staff using AI tools without approval. The inventory step usually finds some; a clear policy and approved alternatives work better than bans.
How do you handle personal data in prompts?
By minimizing what is sent, masking identifiers where possible, and choosing processing locations and terms that fit your obligations.
Can logs themselves become a risk?
Yes, because they contain prompts and answers. Logs get the same access controls and retention limits as the source data.
How often should AI systems be retested?
After every model, prompt, or tool change, and on a regular schedule for high-risk systems.
Can you review a vendor's AI product for us?
Yes. We review data flows, retention, hosting, and controls, and summarize the risks in plain language.
Do small teams need all of this?
No. Controls scale with risk; a small team may need only an approved-tools list, basic logging, and a one-page policy.
How do you test agents that use tools?
With scenarios that try to make the agent misuse tools, leak data, or skip approvals, and by checking every tool call in the trace.
Which framework should we follow?
Many teams use OWASP's LLM Top 10 for technical threats and NIST's AI RMF for governance; they complement each otherSource [1]Source [2].
What's the single most important AI control?
Least privilege: limit what the AI can access and do, and require approval for high-impact actionsSource [3].

Sources

  1. [1]OWASP Top 10 for LLM Applications 2025 (opens in a new tab)OWASP Gen AI Security Project, 2025
  2. [2]AI Risk Management Framework (opens in a new tab)NIST, released 26 January 2023
  3. [3]LLM06:2025 Excessive Agency (opens in a new tab)OWASP Gen AI Security Project, 2025