Skip to main content
SENTRIX / AI-driven in-depth vulnerability assessment Automation + red team Information security disclosure · ISMS-P readiness

AI Penetration Testing
that validates attack paths

AI explores the attack paths. Our red team validates them in practice.
We look for connected risks that lead to unauthorized access, payment abuse or data leaks.

  • Web and API attack paths
  • Access control and payment logic
  • LLMs and AI agents
Who it's for
AI and complex web servicesBefore launch · Feature expansion · Security reviews
Key deliverables
Attack paths and fix guidanceReproduction evidence · Risk rating · Retest
How to start
Consultation on targets and scheduleQuote based on scope · Reply within one business day

Compare scope and deliverables with our other services →

Assessment scope by target

Attack path validation for standard web apps and APIs

AI testing tools and experts validate login, permission and payment flows. Standard web apps and APIs without LLM features can also be assessed.

Synthetic example: a permission flaw that lets one user access another user's orders → reproduction conditions, fix guidance and retest record.

Sample web and API deliverables
Contact us about a web and API assessment →

Security validation of the LLM or AI service itself

We check prompt inputs, access rights to the documents the service retrieves, and how far AI agents can run tools and pass on data. We review the application context, not only the model's responses.

Synthetic example: an attempt to retrieve documents without permission or to run a tool without approval → the conditions that allowed it, observed results and residual risk, all recorded.

Data, tools, test accounts and exclusions are agreed in advance. This service does not guarantee that every bypass will be detected, and it does not certify model safety.

Contact us about an LLM or AI security assessment →

Examples to illustrate deliverables, not real customer results. Service overview and sample reports →

Free assessment for businesses

One web service, one free AI web assessment

One free AI web assessment for one web service run by a registered business. You receive an expert summary report on the AI analysis results. Manual validation of post-login permissions and business logic is scoped separately when we discuss the results. We review your application and then let you know the schedule. Work starts after an NDA is signed and written authorization is given.

Apply for a free assessment
01 / WhyGoing deeper

Beyond the checklist,
down to attack paths.

We check whether individual vulnerabilities could actually lead to a breach.

What automation does

Broad, repeatable checks

  • Collects exposure and configuration information for the agreed targets
  • Runs pattern-based checks and organizes vulnerability candidates
  • Records observations by target and access condition
  • Gathers leads to review from the code and responses provided
What experts confirm

Business context and impact

  • Reviews the reproduction conditions and impact of chained attack paths
  • Designs scenarios that reflect access rights and business rules
  • Separates confirmed results from candidates that need further checking
  • Explains fix guidance and the agreed retest scope
02 / MethodHow we assess

AI explores.
The red team validates.

A four-step process combining AI automation and expert validation

01ExploreAI

Attack surface mapping

Assets · APIs · Data flows

02Find candidatesAI

Vulnerability discovery

Exposure checks · Attack input variation

03Validate in practiceAI + human

Scenario testing

Authorization bypass · Payment logic · Chained attacks

04RemediateHuman

Reporting and retest

Reproduction evidence · Fix guidance · Confirming results

Experts make the final call Candidates found by AI are reviewed against the service context and actual reproduction results. The retest scope and number of retests are set in the contract.

03 / ScopeWhat we look for

Main test areas

From web apps and APIs to AI features, we choose the items that fit your service.

01

Chained attacks

Authentication bypass → Privilege escalation → Data leak

02

Authorization and authentication bypass

Access to other accounts · Admin privilege takeover

03

Payment and business logic

Payment amount tampering · Abuse of order and settlement flows

04

API and server configuration

Exposed admin functions · Missing access controls

05

Prompt injection

Bypassing AI instructions · Injection through external documents

06

Sensitive information disclosure

Other users' data or data from connected systems exposed in responses

07

AI agent misuse

Retrieval data poisoning · Tool permission abuse

08

LLM service protection

Unsafe output · Excessive calls · Cost exhaustion

04 / AudienceWho needs it

Main types of services
that need an in-depth assessment

A

Subject to information security disclosure or ISMS-P

Companies with a deadline for a required assessment under Korea's information security disclosure (an annual public report on a company's security investment and staffing) or ISMS-P (Korea's information security and personal information management certification). Organizations that need to prove real defensive capability, not just pass a formal check.

B

Services with AI or LLM features

SaaS products and apps that have added chatbots, support bots or AI agents. We review the attack surface of the new features.

C

Finance · Healthcare · Commerce

Services built around personal data, payments and transaction logic. We review the potential for data leaks and business damage.

D

Public institutions · Large web services

Organizations that want to build on existing assessment results and further review attack scenarios for permissions and business flows.

05 / TrustSENTRIX track record

A team with hands-on experience
in both automation and attack

In-house scanner + manual penetration testing + AI scenario testing.
We assess from an attacker's point of view and deliver evidence your team can explain to others.
10 years in security Red team operations 2025 AI Red Team Challenge hosted by Korea's Ministry of Food and Drug Safety (MFDS), Top Excellence Award (team member) Vulnerabilities reported to Korea's national database (KVE) In-house automated scanner (CodeScan) OWASP LLM Top 10 Information security disclosure · ISMS-P

These are competition, teaching and vulnerability-report records. They do not imply any partnership, sponsorship or client relationship with the organizations mentioned.

06 / FAQQuestions

Frequently asked questions.

How this differs from SENTRIX ASM

ASM finds your organization's external assets and brand threats, and tracks ownership evidence, owners, fixes and recheck status. AI-driven in-depth vulnerability assessment works from the agreed web apps and APIs, test accounts and attack scenarios. It reproduces risks such as authorization bypass and business logic flaws, and gives you the evidence needed to fix them. If it is unclear which assets you need to manage and who owns them, consider ASM. To confirm the attack paths in a specific service, consider the in-depth assessment. Targets, support and recheck scope are agreed separately for each service.

See the ASM overview and engagement options →

How this differs from an automated vulnerability scanner

Automated scanners are used for repeatable, pattern-based checks and for collecting vulnerability candidates. Paths where business rules and user permissions intersect need to be checked in the context of the service. With the agreed accounts and scenarios, SENTRIX reviews whether each candidate can be reproduced and what its impact is, and reports the supporting evidence, fix guidance and items that need further checking, each kept separate.

What AI does and what experts do

AI supports reconnaissance and asset mapping, test input variation, and review of code and data flows. Experts review business rules, test scenarios, reproduction results and impact. Candidates from automated analysis are kept separate from confirmed results. Tests that could affect operations, and the retest scope, are agreed in advance.

Scope of AI and LLM application assessments

We can assess applications that use models you trained yourselves or external LLM APIs. Along with web and system vulnerabilities, we test the attack surface against the OWASP LLM Top 10, including prompt injection, sensitive information disclosure, RAG data poisoning and misuse of AI agent permissions.

How cost and timeline are set

Cost and timeline depend on the target scope, assessment depth and urgency. If you have a deadline, for example for information security disclosure or ISMS-P, we agree on a priority schedule. Quotes and proposals are free, and we reply within one business day of your inquiry.

How we protect your data and systems during testing

We sign an NDA, use least-privilege access, follow a non-intrusive-first approach and destroy data after the assessment ends. Tests that could affect service availability are agreed in advance, and the vulnerabilities we find are reported only to you.

What to check first to prepare for AI-assisted attacks

In publicly reported AI hacking incidents, AI was used less to create new attack methods than to find existing flaws faster and in greater numbers. The weak points were admin functions and debug pages exposed to the internet, weak authentication, secrets left in apps and repositories, and late patching. Check your external asset inventory, multi-factor authentication, secret cleanup and time to patch first. Then confirm, within the agreed scope, whether issues in post-login permissions and business logic, and in LLM and AI agent permissions, can be reproduced.

Scope the business flows
you need to check now

We confirm the target service, test accounts and the business flows to review. After agreeing on scope, schedule and operational impact, we propose how to validate them. At an early review stage, you can start with the free one-time AI web assessment for businesses.