Agentic penetration testing

Agent speed. Human judgment. Proof on every finding.

RedVault runs authorized penetration tests of your web apps, APIs and AI agents. AI agents do the reconnaissance and testing at machine speed. A RedVault security engineer sets the scope, approves every exploitation step, and reproduces every finding before it reaches you.

Written authorization before any test

Every agent action logged

Every finding reproduced by a person

One retest included

What we test

Two attack surfaces, tested in depth.

Web applications and APIs

Authenticated and unauthenticated testing of web apps and REST or GraphQL APIs. We look hardest at what scanners miss: who can reach what.

  • Broken access control and object-level authorization (IDOR)
  • Authentication, session and token handling
  • Injection and server-side request forgery
  • Business-logic flaws in real user flows
  • API authorization at the object, property and function level

AI and agent applications

Products built on large language models, and agents that call tools, browse, or act for users. We test what the model can be talked into, and what it can reach once it is.

  • Direct and indirect prompt injection, including through documents, web pages and tool output
  • Tool and permission abuse: agents doing more than the user is allowed to
  • Leakage of system prompts, secrets and other users' data
  • Unsafe handling of model output in downstream systems
  • Memory and context poisoning across sessions

Network infrastructure, cloud configuration reviews, social engineering and physical testing are not part of this offer. If you need them, we'll say so on the scoping call.

How an engagement works

Five steps, one gate.

  1. 01Scoping call. Thirty minutes on what's in scope, which environments and test accounts we'll use, and when. You get a fixed quote and a draft rules of engagement.
  2. 02Authorization. You sign the rules of engagement and a written authorization that names every target. If a target is hosted by someone else, you confirm you're allowed to authorize testing of it. No testing starts before this.
  3. 03Agentic testing. Agents map and test the in-scope targets. A RedVault engineer watches the run and approves each exploitation step. If we find something critical, we tell you the same day.
  4. 04Report and readout. A written report with an executive summary and findings you can reproduce, then a call to walk through them.
  5. 05Retest. Once you've fixed the findings, we retest them one time at no extra cost and update the report.

Agentic, with a human in charge

Agents do the legwork. People make the calls.

What the agents do

  • Map the attack surface: routes, endpoints, parameters and roles
  • Generate and run test cases across every role and object you give us
  • Run prompt-injection and tool-abuse chains at a volume a person can't
  • Collect evidence for each candidate finding

What a RedVault engineer does

  • Sets the scope and loads it into the agents' guardrails before the run
  • Approves every exploitation step and every action that writes data
  • Reproduces every finding by hand; anything we can't reproduce is dropped
  • Writes the report and is your single point of contact

Guardrails

  • All agent traffic runs through an intercepting proxy that blocks anything outside the signed scope
  • Every agent action is logged with a timestamp and kept with the engagement record
  • The rate limits you set are enforced on every request
  • One stop command ends all agent activity at once

What you get

Findings you can act on, not a scanner dump.

  • Executive summary. One page: what we tested, what we found, and the overall risk in plain words.
  • Findings. Each one has a severity (CVSS v4.0, adjusted for your business context), an ID and CWE, the exact location (URL, endpoint or prompt path), step-by-step reproduction with request and response evidence, the impact, and a specific fix.
  • Coverage statement. What we tested, with which roles and accounts, and what we didn't test and why.
  • Agent activity summary. How many requests and test cases ran against each target. Full logs on request.
  • Machine-readable findings. A JSON export alongside the PDF, so findings can go straight into your tracker.
  • Retest update. After the retest, the report shows which findings are fixed and which remain.

Example finding

Illustrative · not from a client engagement

HIGH VUL-AI-INJ-INDIRECT CWE-1427

Indirect prompt injection in an uploaded document triggers the send_email tool

POST /api/assistant/chat · tool send_email

Remediation

Require user confirmation for outbound tools, and treat document text as data, not instructions.

Authorization, safety and data

Authorized, scoped, and careful with your data.

  • Written authorization first. We only test systems you own or are entitled to authorize, and only after you sign an authorization that names each target.
  • Staging first, production by agreement. We prefer a staging environment. If we test production, it's with your agreement, in an agreed window, at an agreed rate.
  • Nothing destructive. No denial of service, no social engineering, no deleting or altering your data. We prove access with the least data needed.
  • Dedicated test accounts. You create accounts for each role we test. We never use real users' credentials.
  • AI providers. Our agents use commercial AI models through their APIs, under terms that don't let the provider train on our data. The provider may retain prompts and outputs for a limited time for safety monitoring.
  • Retention. We delete engagement data 90 days after the final report, unless you ask us to delete it sooner.

Read the full methodology and rules of engagement →

FAQ

Common questions.

What does “agentic” mean here?
AI agents do much of the reconnaissance and testing: they map your app, generate test cases and run them. A RedVault engineer directs the work, approves every exploitation step, and confirms every finding before it's reported.
Is this an automated scan?
No. Scanners report what matches a pattern. Our agents test real user flows and roles, and a person reproduces each finding. You won't get unverified results.
Will you test production?
Only if you agree to it in writing, in an agreed window and at an agreed rate. We prefer staging.
How long does an engagement take?
It depends on scope. The quote states the testing window, and we agree it on the scoping call.
What do you need from us?
A signed authorization, test accounts for each role, API documentation or an OpenAPI spec if you have one, and a contact we can reach during testing.
Can you test an app built on a third-party AI model?
Yes. We test your application and how it uses the model. We don't attack the model provider itself.
Will you sign an NDA?
Yes. We sign a mutual NDA before you share scoping details.
Who will I work with?
One RedVault security engineer, from the scoping call through the retest.

Scope a test in 30 minutes.

Tell us what you want tested. We'll reply within one business day with times for a scoping call.

RedVault Security Inc. tests only systems the client is authorized to have tested, under a signed authorization and rules of engagement.