Prompt injection tied to data or action impact
Assess how instructions from users, retrieved content, files, and tool responses can influence the workflow. Connect observed behavior to data exposure or unauthorized action.
LLM applications · RAG · Agents · MCP
AI Application & Agent Security Assessment
Follow the data. Test the permissions. Validate the actions.
For engineering, AI, and security leads responsible for production systems that read sensitive information or change business records. Scope the application, retrieval, agents, and tools around the model.
Scope & deliverables
What the engagement covers.
RAG and cross-tenant access
Examine retrieval authorization, document ownership, index filters, metadata, and tenant boundaries. Use representative identities and synthetic customer data.
Agent identity and excessive agency
Build an agent identity and tool-permission map: acting identity, delegated scopes, accessible data, allowed actions, and approval requirements. Test whether a workflow can cross those agreed boundaries.
MCP and tool calling
Scope hosts, clients, servers, tools, resources, and transports. Assess authorization boundaries, tool inputs and responses, server trust, and permissions across downstream integrations.
Memory and context manipulation
Examine whether untrusted context, persistent memory, or shared conversation state can affect later decisions, expose data, or cross user and tenant boundaries.
Models, providers, and data handling
Map model/provider boundaries, routing, sensitive-data flow, output handling, and relevant environment controls. Record tested versions and limitations.
Methodology references
Framework-informed. Product-specific.
We map relevant risks from OWASP LLM guidance , OWASP Agentic Applications , and MCP security guidance to your architecture and agreed test objectives. These references inform coverage; they are not certifications or guarantees.
Architecture and evidence
Connect AI behavior to a business consequence.
Follow an end-to-end abuse path
Trace untrusted input through retrieval, memory, agent identity, and tool execution. Assess whether it can expose another tenant’s data, bypass approval, or perform an unauthorized action within the agreed test environment.
Keep the assessment boundary clear
This offer covers production applications and integrations. Foundation-model weights, training infrastructure, workplace-AI rollout, and specialist hardware require separate capability and scope. Review the illustrative evidence structure.
Commercial scope
Know what you are commissioning.
Starting engagement range: $12,000–$20,000 USD, depending on scope. This is a starting-scope range, not a cap. Workflow count, agents, MCP servers, tools, retrieval sources, permissions, and tenants determine the quote. Combined SaaS and AI testing is itemized.
Before work begins, agree delivery dates, authorized access, exclusions, data handling, and payment terms. Review our methodology and report expectations.
Practical answers
Before we start.
What is delivered?
A scoped threat model, role/tenant/workflow coverage matrix, agent identity and tool-permission map, validated findings, buyer evidence summary, remediation workshop, and one scoped retest. The proposal defines workshop time; the report records versions, configurations, and reproducibility limits.
Is the OWASP guidance a certification?
No. We use relevant OWASP LLM and agentic risk guidance alongside product-specific test objectives. A checklist does not establish complete coverage or certify that a system is secure.
How do we handle changes after testing?
Record changes to models, tools, data sources, memory, or permissions against the tested baseline. After reviewing the baseline results and upcoming changes, choose targeted follow-up or a Continuous AI Assurance plan where repeat testing is justified.
Does this guarantee safe model behavior?
No assessment can guarantee every future output. Evidence applies to the workflows, versions, permissions, and conditions tested. The report states coverage, uncertainty, and exclusions.
Prepare for the engagement
Relevant buyer guides.
AI Application & Agent Security: Scope & Cost
Practical guidance for scope, evidence, and the next security decision.
MCP Security Testing: Scope & Risks
Practical guidance for scope, evidence, and the next security decision.
Continuous AI Assurance: Annual Testing Plans
Practical guidance for scope, evidence, and the next security decision.
Start with a scoped conversation
Bring the requirement.
Bring the requirement.
We’ll define the right scope.
Share your product, buyer requirement, and target date. We’ll confirm the engagement, deliverables, and commercial terms before work begins.