Skip to preparation guide
← DefendML home

Before your AI security assessment: prepare the boundary, not just the chatbot

What should your application never let an AI assistant read or do—and what legitimate work must still succeed?

For Product Security and AI engineering teams, those two questions make a pre-deployment review more concrete. DefendML is an offensive AI red-team testing service: we test authorized scope and retain evidence for review. We do not provide runtime protection or apply fixes to your application automatically.

Bring a short, non-secret description

  • Your goal: the release decision or security question you need evidence for.
  • Your application: what it does, how users interact with it and its interface type. Do not send credentials or secret-bearing endpoint URLs.
  • Your authority: who owns the application and who can approve the exact test scope, environment and window, including any third-party restrictions.
  • Your technical owner: who can explain the setup, implement changes and coordinate a retest.
  • Your boundary: one example of a prohibited read or action, and one legitimate task that must continue to work. Describe them without actual customer records.
  • Your deliverable needs: who will review the evidence and any privacy, procurement, export or scheduling requirements to discuss—not assumed commitments.

Confirm fit before connecting anything

DefendML supports testing through supported HTTP AI application interfaces; compatibility must be reviewed for your request format and intended observations. A working URL or a model's refusal does not prove that downstream access was denied.

For an initial founder-supervised RAG assessment, discuss an owned staging application with synthetic documents, a prohibited cross-organization read and a legitimate same-organization control. RAG means an assistant retrieves documents to help answer a question. This scoped setup requires a compatible adapter and reviewed observations from the application; it is not a general RAG attack pack or a promise that every chatbot can be assessed unchanged.

Where MCP is relevant, current native testing is limited to configured resource boundaries over stateless Streamable HTTP. Do not assume comprehensive OAuth, tool, prompt or stateful-session coverage. Agree the actual interface and scope first.

Use synthetic data and agree the stopping conditions

Prepare staging data that does not contain real customer information. Before execution, agree testing authority, permitted destinations, access limits, stop conditions and cleanup responsibilities. Required credentials belong only in agreed supported secure setup controls—not inquiry email or ordinary chat. Confirm the authorized scope through Review & Run before running an assessment.

Know what the evidence can—and cannot—establish

Review findings and advisory Fix Plan guidance against the recorded scope. Your team implements the correction; a comparable full-target retest evaluates the changed application. Check legitimate behavior separately, not just the attack. A technical error or missing observation is not proof that an action was blocked. “Verified resolved” applies only when the relevant evidence requirements are met.

Our controlled RAG case study shows a finding-to-retest workflow using an owned application, synthetic data and deterministic execution, with no language model tested. It is not a customer result or evidence of general real-model effectiveness.

In-portal evidence and formal exports are different deliverables. Complete PDF Reports and summary CSV/JSON exports require an organization export entitlement and explicit inclusion in the agreed assessment. No price, turnaround, certification, complete coverage or safety guarantee is implied.

Take the next step

Share a brief non-secret summary of your application, technical owner, proposed boundary and legitimate task. Scope and setup are reviewed before testing; an inquiry is not automatic approval or a commitment to an assessment.

Discuss a scoped assessment