AI penetration testing

    AI penetration testing for LLM applications and AI agents

    Find out what your AI features will do for an attacker, before one asks.

    BlueRadius provides AI penetration testing for companies shipping chatbots, copilots, retrieval over internal data and AI agents that act on a user's behalf. We attack the model, the data it can read, the tools it can call and the application around it, then show your engineers how to close each path.

    AI features fail in ways a standard pen test is not built to find. A model that follows instructions hidden in a document, or an agent with broader access than the user it serves, can expose data or take actions no one approved. You get a report your customers and auditors will accept and a retest once the fixes ship.

    When companies need an AI pen test

    • You are launching an AI feature to customers: a chat assistant, a copilot, AI search or summarization over their data.
    • You are giving an agent real access: email, a CRM, a ticketing system, code execution or internal APIs.
    • An enterprise customer's security review asks how you tested your AI features against prompt injection and data leakage.
    • A SOC 2 or ISO 27001 audit where the auditor wants evidence that a new AI system was tested before release.
    • Your board or AI governance committee wants independent assurance before an internal agent goes live.

    What we test

    Prompt injection, direct and indirect

    We try to override your instructions from the chat box, and then from every place the model reads content it did not write: uploaded files, retrieved documents, emails, tickets, web pages and tool results. Indirect injection is where most real damage comes from, so we plant payloads in the data sources your AI actually consumes.

    Data leakage through retrieval

    If your AI searches documents or records to answer questions, we test whether one user can get it to retrieve another user's or another customer's data. Retrieval often skips the permission checks the rest of your application enforces, and multi-tenant SaaS is where this matters most.

    System prompts, secrets and configuration

    We check whether the model can be talked into revealing its system prompt, API keys, internal URLs or connection details, and whether anything sensitive was placed in the prompt that should have lived behind an access check instead.

    Tools, agents and excessive agency

    For agents and function calling, we map every tool the model can reach and test whether an attacker can make it call them: sending messages, changing records, running code or moving data out. We check whose identity and permissions the agent uses, whether it can act beyond what the current user is allowed to do, and whether approval steps for risky actions can be skipped.

    Output handling

    Model output is untrusted input to the rest of your system. We test whether it can carry cross-site scripting into your interface, injection into queries or commands, or links and markdown that quietly send data to an attacker's server.

    Supply chain and integrations

    Plugins, MCP servers, third-party connectors, open-weight models and the libraries around them. We review what each one can access and test whether a compromised or malicious component could reach your data.

    The application and APIs around the model

    Authentication, authorization, rate limits and the APIs your AI feature calls, tested the way a web application penetration test would. Many AI findings end here: an endpoint that trusted the model's request because it came from inside.

    How an engagement runs

    1. Scoping. We agree the AI features, models, data sources, tools, roles and environments in scope, and who will read the report.
    2. Design review. A walkthrough of the system prompt, retrieval setup, tool definitions and permissions, so we test the real architecture rather than guess at it.
    3. Testing. Manual attack chains supported by automated payload runs. Because model responses vary, we repeat attacks to confirm how reliably each one works. Critical findings are reported the same day.
    4. Reporting. An executive summary and each finding with severity, reproduction steps, evidence, its OWASP LLM Top 10 and MITRE ATLAS references, and a fix your engineers can apply.
    5. Retest. Once fixes are deployed, we verify them and issue an updated report.

    Why BlueRadius for AI security testing

    We research how AI agents fail in the real world. Our AI Agent Security Report and public AI agent incident tracker document the failure patterns we test for, from injected instructions to agents acting with more access than anyone intended.

    Testing is one part of running AI safely. Our AI governance practice sets the policies, inventory and approval process around new AI systems, a virtual CISO can own the remediation plan, and if the test is for a compliance program, our SOC 2 consultants make sure the report answers what your auditor will ask.

    Frequently asked questions

    What is AI penetration testing?

    A hands-on security test of an application that uses a large language model: a chatbot, a copilot feature, a retrieval (RAG) search over your documents, or an agent that can call tools and take actions. The tester tries to make the AI do what it should not, such as reveal another user's data, ignore its instructions, leak secrets, or use its tools against you, and shows what an attacker could actually reach.

    How is it different from a normal web application pen test?

    A web app test looks for flaws in code and access control. An AI pen test adds the model itself as an attack surface: the model follows instructions written in plain language, including instructions hidden in documents, emails or web pages it reads. That makes prompt injection, data leakage through retrieval, and agents with too much access the main concerns. We test both layers, because most serious AI findings end in a conventional failure, such as an API that trusted whatever the model sent it.

    What is prompt injection?

    Getting a model to follow an attacker's instructions instead of yours. Direct prompt injection comes from the user typing into the chat. Indirect prompt injection is hidden in content the model reads on someone's behalf, such as a support ticket, a shared document, an email or a web page. Indirect injection is the bigger risk for agents, because the person affected never sees the malicious text.

    What standard do you test against?

    We use the OWASP Top 10 for LLM Applications (2025 edition) for coverage and MITRE ATLAS to describe attacker techniques, alongside the OWASP Web Security Testing Guide and API Security Top 10 for the application and APIs around the model. The report maps each finding to those references so your engineers and auditors can see where it fits.

    Do you need access to the model itself?

    No. Most companies use a hosted model from a provider such as OpenAI, Anthropic or Google, and we test your application, not the provider's model. We need test accounts for each role, a description of what the AI feature is allowed to do, and ideally the system prompt and tool definitions, so we can spend the time on attacks instead of guessing at the design.

    Can you test AI agents that take actions?

    Yes, and that is where testing matters most. When an agent can send email, update records, run code or call internal APIs, a successful injection becomes an action rather than a bad answer. We test which tools the agent can reach, whose permissions it acts with, whether risky actions need human approval, and whether those controls hold up under attack. Testing that changes data runs in a staging environment.

    Is retesting included?

    Yes. Once your team has fixed the findings, we retest them and issue an updated report showing what was resolved. Because model behavior can vary between runs, we retest each injection finding several times rather than once.

    Scope your AI pen test

    Tell us what the AI feature does, which model and data it uses, what tools or actions it can take, and who needs to see the report, and we will come back with a scope and a quote. Book a 30-minute scoping call, or see our full range of penetration testing services.

    Two ways to work with us

    Strategy, and the platform to run it, from one team.

    Services

    Fractional security leadership

    Virtual CISO leadership, managed security, and audit-ready compliance, run by senior practitioners who have done it at Fortune 100 scale.

    Explore services

    Platform

    Radius360

    The security program platform we build on: it turns your signals into decisions with the evidence attached, across 30 frameworks.

    Explore Radius360

    Not sure where you stand?

    Book a free consultation