AI penetration testing | LLM pentest

Prices starting at $8,999.

AI penetration testing is a manual security assessment of AI-powered applications, LLM integrations, agents, and supporting infrastructure. Researchers test whether malicious prompts, documents, retrieved content, or model outputs can expose data, bypass access controls, or trigger unauthorized actions.

Blaze's penetration testing services  combine application security testing with AI-specific attack scenarios. We assess trust boundaries across users, tenants, retrieval pipelines, tools, and integrations, delivering validated findings and practical fixes.

Blaze is a CREST-accredited company. Our researchers hold certifications including OSCP, OSWE, OSCE, and CRTO.

Assessments draw on the OWASP Top 10 for LLM Applications (2025), MITRE ATLAS, and relevant application-security guidance. Findings can inform risk-management activities under the NIST AI RMF.

Explore our AI and LLM penetration testing approach .

Testing is tailored to your architecture, permissions, deployment, and available access. Scope can include:

• LLM applications: Chatbots, copilots, and generative AI features
• Prompt injection: Direct, indirect, multi-turn, and multimodal attacks
• Guardrail bypass: Jailbreak resilience and application safeguard testing
• RAG pipelines: Retrieval manipulation, vector-store permissions, tenant isolation, and data leakage
• Agentic AI: Unsafe tool calls, excessive permissions, unauthorized actions, and MCP integrations
• Output handling: Injection, data exposure, and unintended execution through model outputs
• Custom models: Data extraction, model inversion, and membership inference where access permits
• Supply chains: Model dependencies, adapters, data provenance, and poisoning risks
• Consumption controls: Resource abuse, excessive cost, and service disruption
• Supporting applications: Web, API, mobile, authentication, and authorization testing

We assess your deployments and integrations using Amazon Bedrock, Amazon SageMaker, OpenAI, Anthropic Claude, Azure OpenAI, and Google Vertex AI, where included in scope. Testing targets your authorized assets, rather than providers' underlying infrastructure.

Researchers use Garak, PyRIT, Promptfoo, custom tooling, and manual investigation. Automated results are validated before reporting.

Assessment effort typically ranges from 8 to 25 person-days, depending on system count, agent complexity, and tool surface. Testing dates and delivery are agreed during scoping.

You receive:

• Executive summary explaining key risks and business impact
• Validated findings with reproducible prompts, payloads, and attack transcripts
• Severity ratings, remediation priorities, and practical recommendations
• Relevant mappings to OWASP LLM risks and MITRE ATLAS techniques
• Recommendations for permissions, trust boundaries, output handling, and guardrails
• Signed attestation letter documenting scope and assessment completion
• Free re-test within 90 days

Final reports are delivered within five business days of assessment completion.

Where scope and requirements overlap, findings may support SOC 2, ISO 27001, ISO/IEC 42001, and other assurance activities. They can inform AI risk assessments and technical documentation without replacing separate compliance obligations.

Prices start at $8,999, with discounts for early-stage AI startups. Final pricing depends on scope and complexity.

Talk to an expert about your AI pentest .

Email:  sales@blazeinfosec.com 

Phone: +1 347 892 4783 (US/Canada)

Phone: +351 222 081 647 (Europe/international)

Blaze is a CREST-accredited, ISO 27001 and ISO 9001 certified company, with worldwide professional liability (E&O) coverage of $5,000,000 through Hiscox.