Executive risk summary
Business implications, demonstrated exposure, and recommended priorities.
Offensive Security
AI Security Testing
Your AI can take action.
Make sure it stays within bounds.
AI penetration testing and red teaming for applications, RAG workflows, and agents. Evaluate data access, tool permissions, and safeguards within an agreed scope.
IRON FIST LABS / 16Offensive Security
01 / THE CHALLENGE
Helpful output is not proof of a secure workflow.
An AI workflow can follow the wrong instructions with legitimate credentials. Prompts alone do not enforce authorization. Examine the full path from untrusted content to data access and tool execution.
02 / WHAT YOU RECEIVE
Expertise you can
put to work.
A clear handoff for the people making decisions and the people delivering the work.
Evidence-backed findings
Reproducible scenarios where feasible, affected components, observed behavior, and testing limitations.
Practical remediation guidance
Recommendations for permissions, data isolation, approval checks, and supporting application controls.
Results walkthrough and follow-up plan
Review findings with your team and define any retesting scope and window in the proposal.
Your proposal confirms the deliverables, scope, responsibilities, and schedule for your engagement.
03 / HOW WE WORK
A clear path.
From start to next steps.
Map the workflow
Agree the use case, models, data sources, tools, user roles, environments, and boundaries.
Design the scenarios
Select relevant misuse scenarios, safe test data, approvals, and stop conditions.
Test the boundaries
Evaluate prompt injection, unauthorized data access, tool misuse, and conventional application weaknesses within scope.
Explain and validate
Discuss evidence and priorities. Agree how remediation will be verified and what remains untested.
CHOOSE THE RIGHT DEPTH
Three ways to start.
Architecture review
Examine data flows, permissions, trust boundaries, and approval controls before selecting changes or deeper tests.
AI penetration testing
Test an agreed application or workflow for exploitable weaknesses across AI behavior and supporting controls.
Adversarial scenario testing
Evaluate realistic misuse scenarios across multiple steps, including tool actions, memory, and retrieved content.
Our team includes an OffSec AI Red Teamer (OSAI) certified professional. Each engagement is tailored to the agreed workflow and business objective.
A STRONG START
Bring the context.
We’ll shape the plan.
You do not need every answer before getting in touch. These details help us scope the work.
A useful starting point
A workflow diagram, tool inventory, data sources, user roles, and your key concerns. An early draft is enough to begin.
The right people
An application owner, an AI or engineering lead, and the person responsible for data access and security.
What shapes the engagement
The number of workflows, integrations, user roles, environments, and the depth of adversarial scenarios shape the work. We agree timing after reviewing these factors.
SEE THE SHAPE OF THE WORK
From observation
to next action.
A short example of how we make findings useful. The final format and depth depend on your agreed engagement.
Discuss your deliverablesILLUSTRATIVE EXAMPLE01 / BRIEF
An agent attempts an unapproved data transfer
Fictional scenario. Not a client result or a completed assessment.
- Focus
- An assistant with access to customer records and an email tool.
- Observation
- In this fictional scenario, instructions planted in a document cause the agent to request a transfer to an external recipient. The tool lacks an independent recipient check.
- Recommended next step
- Enforce destination and record-level authorization outside the model, then retest using safe data.
CONTEXT → OWNERSHIP → ACTION
04 / BEFORE WE BEGIN
A little clarity.
A better start.
What can you test?
The scope can include AI applications and chatbots, RAG and knowledge access, agent tools, persistent memory, supporting APIs, and relevant infrastructure.
Do we need a production deployment?
No. A staging environment with representative permissions, integrations, and safe test data can support testing. We agree what the environment can demonstrate.
Is this different from a normal penetration test?
AI testing adds adversarial scenarios involving retrieved content, model behavior, and agent actions. Authentication, tenant isolation, APIs, and other conventional controls remain part of the agreed scope.
Does passing make our AI safe?
No. Results reflect the scenarios, configuration, access, and time tested. Model or workflow changes can alter behavior, so limitations and residual risks are documented.
Is retesting included?
Retesting availability, covered findings, and the window are specified in your proposal. We agree follow-up before the engagement begins.
What AI security experience does your team have?
Our team includes an OffSec AI Red Teamer (OSAI) certified professional. We scope the work around your application and business risks.
LET’S DEFINE THE RIGHT ENGAGEMENT
Start a conversation.
Make your next
move clearer.
Tell us what prompted your search, what matters to your business, and any deadline. We’ll help define the right scope.
What happens next
- We review your priorities.
- We discuss the right scope with you.
- You receive a clear proposal before work begins.
START A CONVERSATION
Tell us what you need.
A clearer picture. A practical next step.
Tell us what you’re working on.
We’ll use your details to respond to your request. Please leave out passwords and sensitive incident evidence. Privacy policy.
Prefer email?info@ironfistlabs.com