Who it is for
Companies deploying customer-facing or internal AI features, security teams asked to sign off on an AI project, and leaders who need an AI usage policy that people will actually follow.
LLM applications introduce failure modes traditional security reviews miss: prompt injection, data exfiltration through retrieval, over-permissioned agents, and leaked secrets in context windows. We find those issues before attackers or auditors do.
Deliverables
LLM threat modeling
A structured review against the OWASP Top 10 for LLM applications, mapped to your architecture and data flows.
Adversarial testing
Prompt injection, jailbreak, data exfiltration, and tool-abuse testing against your real application, with reproducible findings.
Guardrails and output controls
Input filtering, output validation, PII detection, and policy enforcement layers that fit your stack.
RAG data governance
Access control at retrieval time, PII handling, and audit trails so an assistant never surfaces what a user should not see.
Agent permissioning
Least-privilege tool access, sandboxing, approval gates, and spend limits for autonomous agents.
AI usage policy and governance
A practical policy, vendor review checklist, and training for staff using AI tools day to day.
Engagement approach
Scope
Inventory AI features, vendors, models, and data sources in use or planned.
Assess
Threat model plus hands-on testing of the highest-risk paths.
Remediate
Prioritized fixes, implemented with your team or by ours.
Govern
Policy, monitoring, and a repeatable review process for new AI features.
What you walk away with
- Documented, prioritized AI risk register with fixes underway
- Security sign-off that unblocks AI launches
- Guardrails that catch injection and leakage in production
- A governance process that keeps pace with new AI tools
- OWASP LLM Top 10
- NIST AI RMF
- Model Armor
- Cloud DLP
- Sensitive Data Protection
- Secret Manager
Common questions
Is this a penetration test?
It includes adversarial testing focused on AI-specific issues, but it is broader: architecture, data governance, and policy. We coordinate with your existing pen-test vendor when you have one.
We only use ChatGPT and Copilot internally. Do we need this?
A lighter engagement fits: an AI usage policy, vendor data-handling review, and staff guidance. It prevents the common problem of confidential data pasted into consumer tools.
Can you help with compliance frameworks?
Yes. We map findings to SOC 2, ISO 27001, HIPAA, and the NIST AI Risk Management Framework so the same work feeds your audits.
Often combined with
AI & Agentic AI Solutions
RAG assistants, LLM applications, and autonomous agents built on your data and wired into your real workflows.
Learn moreCloud Security Audits
Configuration, IAM, network, and detection reviews across Google Cloud, AWS, and Azure with a prioritized remediation plan.
Learn moreFractional CTO & Technology Leadership
Executive-level technology strategy, architecture governance, vendor management, and team leadership, sized to your business.
Learn moreTalk to us about AI Security
A 30-minute call is enough to tell whether this is the right engagement and what it would take.