Service
AI Application Security Assessment and Red Teaming
Who it is for
Product and engineering teams that have shipped, or are about to ship, LLM features or agents that touch customer data or production systems.
When clients call
- An agent connected to production data or tools
- A customer asking for AI-specific test evidence before signing
- A RAG index built from data nobody classified
- A model or provider upgrade that changed system behaviour
Scope
- Direct and indirect prompt injection, jailbreak and data-extraction attempts
- RAG index content and access control
- Tool, function and MCP server permissions, including chained actions
- Human approval gates and audit logging
- Adversarial testing against an agreed objective, with retest of fixed findings
Method and standards
- OWASP Top 10 for LLM Applications 2025
- OWASP Agentic AI threats
- MITRE ATLAS
- ISO/IEC 42001:2023
Deliverables
- Findings register mapped to OWASP and ATLAS
- Agent permission matrix
- Attack technique log with reproduction steps
- Retest letter
Engagement arrangements
Dates, responsibilities and any ongoing review cadence are agreed in the written scope. Retesting is included where stated in that scope.
What happens after the scoping call
Written scope and fixed-fee proposal within one business day of your scoping call. The scope names the standards applied, the deliverables, the dates and the people doing the work.
Questions
Do you need access to the model provider?+
No. We test your application, its retrieval, tools and permissions. The model provider is out of scope unless you self-host.
What do you need from us for adversarial testing?+
A test environment or a production system you approve for testing, credentials for the accounts in scope, and a named contact who can stop the test immediately.
How is the fee set?+
The written proposal defines the fixed fee, responsibilities and deliverables. Any change to scope is agreed in writing before additional work begins. Ongoing advisory includes an agreed review cadence and availability.
How is our evidence handled?+
Evidence is held under NDA, encrypted at rest and in transit, restricted to the named engagement team and destroyed on the schedule stated in the scope. It is never used to train any model.
Tell us what changed.
Thirty minutes, no obligation. Written scope and fixed-fee proposal within one business day of your scoping call.
Book a Strategic AI Readiness Assessment