SolidLab

Comprehensive security assessment (penetration testing)
of AI-powered applications

Adopting generative AI and LLM agents fundamentally changes application architecture and requires a new approach to security testing. Traditional methods and tools fall short for AI systems. We find and prove AI-specific risks where others miss them.

Why AI systems need specialized assessment

AI components differ from traditional applications in three key ways, each of which creates new security challenges.

Probabilistic behavior

LLMs can return different responses to the same prompt: output is nondeterministic and can vary for identical inputs. This makes typical deterministic tests inapplicable and calls for adaptive testing methods that account for the probabilistic nature of AI.

AI-specific threat classes

Prompt injection, jailbreaking, data disclosure, integration errors, and excessive agent permissions are threats specific to AI systems. They require specialized methods that go beyond the classic OWASP Top 10 for web applications.

Expanded attack surface

An AI agent becomes an autonomous actor with its own privileges. Multi-agent systems add further exposure: every new tool can become an attack point, and every agent connected to internal services can become a privilege escalation point. The attack surface expands significantly and cannot be fully covered by classic vulnerability scanners or protected by a WAF.

Real threats to AI applications

Each of these threats can lead to data compromise, business disruption, or reputational damage.

Prompt injection

Attackers can take control of an AI agent with specially crafted prompts and gain access to data and the ability to act on behalf of the system.

Sensitive information disclosure

An LLM may disclose sensitive information in its responses: user data, internal documents, API keys, or infrastructure details, including the contents of RAG databases.

Business rule bypass

The agent’s context may not account for every aspect of the underlying business logic. Attackers can exploit this to bypass restrictions even without specialized payloads (such as those used in prompt injection attacks).

Excessive agency

AI agents with excessive access to tools and services open the door to large-scale attacks, including using integrated tools to reach internal infrastructure and to launch spam and phishing campaigns.

Unbounded consumption

Targeted overload of system resources through AI components.

Who needs an AI security assessment

Fintech and banking

AI chatbots, scoring, and fraud prevention — protection against manipulation and customer data leaks

SaaS and product companies

AI assistants, content generation, and automation — protection against functionality abuse

E-commerce

AI recommendations and support chatbots — preventing customer data leaks

Healthcare

AI diagnostics and medical data processing — the strictest confidentiality requirements

Government

AI for handling inquiries and analytics — protecting high-value information

Telecommunications and industrial sectors

AI for network and infrastructure management — security of critical systems

What we assess

Our methodology covers every layer of AI-enabled applications, from the user interface to internal agent integrations. The checklists below are indicative, and the exact scope depends on the target system. We also assess classic ML systems based on the OWASP Top 10 for Machine Learning, covering the entire lifecycle from data preparation to production operation.

Architecture

Architecture and attack surface analysis

A comprehensive review from the user interface and API to specific agents in multi-agent systems, including LLMs, integrations, and security mechanisms.

  • Identifying agent components and entry points for direct and indirect prompt injection
  • Understanding the agent’s business logic: goals, restrictions, and tools used
  • Analyzing data flows between users, LLMs, and tools
  • Identifying implemented security mechanisms
  • Identifying critical paths in multi-agent systems
OWASP LLM

AI-specific vulnerability analysis

Testing follows OWASP guidance and methodologies for AI systems, including the OWASP Top 10 for LLM Applications, OWASP Top 10 for Agentic Applications, and OWASP MCP Top 10, as well as MITRE ATLAS and Google SAIF.

  • Direct and indirect prompt injection
  • Improper output handling
  • Sensitive information disclosure
  • Unbounded consumption and other risks
Integrations

Connected integrations and access rights analysis

Review of connected integrations (resources, tools, RAG systems, and other agents) and the privileges agents hold.

  • Audit of access rights to tools and APIs, including the OWASP API Security Top 10
  • Testing APIs for known classes of web vulnerabilities and for the ability to craft payloads using LLMs
  • RAG system analysis for AI-specific vulnerabilities
  • Authorization checks for calls to external services
  • Horizontal and vertical privilege escalation analysis
  • Session isolation testing
  • Checking for extraction of system prompts and configuration
  • Analysis of architecture and agent hierarchy in multi-agent systems
  • Attempting to exploit trust chains between agents in multi-agent systems
Security controls

Analysis of implemented security mechanisms

Analysis of AI system components for hidden capabilities, potentially dangerous actions, or other excessive permissions.

  • Jailbreaking and content filter bypass
  • Agent restriction analysis
  • Creating and testing agent abuse scenarios
  • Testing scenarios that bypass business logic restrictions, including request chains
Business logic

GenAI functionality abuse

Searching for ways to abuse implemented AI functionality to subvert the application’s business logic.

  • Manipulating agent behavior using system specifics and weaknesses discovered during previous assessment stages
  • Exploiting logical errors in agent chains

Assessment stages

A structured process ensuring completeness and transparency at every step.

01

Scope definition

Agreeing on scope, identifying AI components, and reviewing system documentation and architecture.

02

Architecture analysis

Defining the attack surface: LLMs, agents, integrations, APIs, user interfaces, and data flows.

03

Implementation analysis

Hands-on vulnerability discovery following OWASP guidance and methodologies for AI systems, MITRE ATLAS, and Google SAIF; integration testing; and verification of security controls.

04

Vulnerability demonstration and risk analysis

Demonstrating the real-world consequences of identified vulnerabilities and assessing their severity and business impact.

05

Report and recommendations

Preparing a detailed report with technical details, severity ratings, and specific remediation recommendations.

What you get

A detailed report on identified risks, findings, and their severity, plus recommendations for remediation and improving your overall security posture.

Executive summary

A concise overview of results and key risks, plus an overall rating of the AI system’s security posture to support high-level decisions.

Technical report

A detailed description of each vulnerability: root cause, steps to reproduce, proof of concept (PoC), and a severity rating based on CVSS and AI-specific factors.

Remediation recommendations

Specific, actionable instructions for developers and DevOps teams to fix each vulnerability and strengthen overall security.

✓Retest after remediation
✓Remediation consulting for identified issues

Why SolidLab

Our group of companies has been in the cybersecurity market for more than 15 years. The team conducts research at the Faculty of Computational Mathematics and Cybernetics of Moscow State University, and its results have been presented at leading conferences: OWASP AppSec, DEF CON, Black Hat, Hack in the Box, Positive Hack Days, OffZone, and ZeroNights.

Research expertise

Key team members conduct security research, publish, and speak at leading conferences.

Proven results

Numerous published CVEs in Apple products, Google Chrome, VMware vCenter, and MySQL2. Over 100 accepted reports in bug bounty programs from leading global companies.

Proprietary technologies

Our own tools and expertise let us analyze complex applications and APIs and intelligently validate results.

Confidentiality

An NDA is signed before work begins. All terms of the agreement are strictly followed. All information is securely deleted after project completion.

Industry recognition

Acknowledgments from Alibaba, Amazon, Apple, Google, IBM, PlayStation, Mail.ru, and other companies for responsible vulnerability disclosure.

Critical vulnerabilities discovered by our team

Apple

CVE-2025-24192

Google Chrome

CVE-2023-5480 · CVE-2024-10229 · CVE-2025-4664

MySQL2 for Node.js

CVE-2024-21507 · CVE-2024-21508 · CVE-2024-21509

Questions and answers

How is an AI system security assessment different from a standard web application assessment or a penetration test?

Assessing AI systems means accounting for the nondeterministic behavior of their components: an LLM can return different responses to the same input. An assessment may include testing for known classes of web application vulnerabilities, such as cross-site scripting (XSS) and SQL injection, but it focuses primarily on AI-specific vulnerabilities, LLM restriction mechanisms, AI agent permissions, and ways to bypass business logic through AI components. This requires specialized expertise and up-to-date, domain-specific methodologies.

What types of AI systems do you test?

We analyze a wide range of systems: web applications integrated with LLMs (ChatGPT, Claude, Gemini, etc.), multi-agent systems, RAG systems, AI chatbots and assistants, automation systems based on AI agents, and custom solutions based on open models.

Do you need access to source code?

We can perform the assessment as black-box, gray-box, or white-box testing. For maximum coverage, we recommend gray-box or white-box testing. Even with black-box testing, however, we identify critical vulnerabilities at the user interface and external API levels.

How do you ensure the security of our data?

An NDA is signed before work begins. We guarantee full protection of the data you provide, strict compliance with the confidentiality agreement, and secure deletion of all information after the project is completed. Testing can be performed in an isolated environment.

What happens after the report is delivered?

We provide remediation consulting for identified vulnerabilities and retest after fixes are applied to confirm they are effective. If needed, we can also support you throughout the remediation process.

Find vulnerabilities in your AI application before attackers do

Contact us to discuss the security assessment of your AI system. We will estimate the scope and propose an optimal testing plan.