AI / LLM Pentest · Specialty
AI and LLM penetration testing
Is your chatbot, copilot or agent already in production? We assess the whole system — model, prompts, knowledge base, connected tools and integrations — to show what an attacker could do and how to fix it.
Request a quote See how it worksTypical timeline: 5–12 days
Why AI needs a dedicated pentest
In LLM systems, natural language becomes control input, and every tool connected to the model is a path to data and actions. Web scanners do not assess that behaviour. An AI pentest answers concrete questions: can the system leak other customers' data? Can it take actions the user could not? Can it run up uncontrolled cost?
What we assess
- Instruction manipulation by users and by external content the system reads.
- Data leakage from context, the knowledge base (RAG) and across users and tenants.
- Agents and tools: permissions, human confirmation and limits on actions the model can take.
- Model output used in pages, queries or integrations without validation.
- Cost and availability: usage limits and abusive consumption.
- Supply chain: third-party models, plugins and dependencies.
Coverage is mapped to the OWASP Top 10 for LLM Applications, MITRE ATLAS and the NIST AI RMF — vocabulary product, security and audit teams share.
Who it is for
- Companies with a customer-facing support or sales chatbot.
- SaaS products that shipped an AI copilot or AI search over customer data.
- Teams that connected agents to internal systems, email, CRM or payments.
- Companies that must evidence AI security to customers, auditors or the EU AI Act.
Typical AI system risks
- An assistant revealing other customers' data held in the knowledge base.
- An agent taking sensitive actions (email, refunds, exports) without confirmation.
- External content read by the system changing its behaviour.
- Model output inserted into pages or queries without validation.
- Unbounded usage driving high token cost.
A pentest that holds up in audits
The report includes scope, dates, methodology, justified severity and retest evidence — what auditors and customers look for under SOC 2 (CC4.1, CC7.1), ISO 27001 (A.8.8, A.8.29), PCI DSS 4.0 (requirement 11.4), HIPAA, GDPR (Art. 32), NIS2 and DORA. We also issue an attestation letter you can share with customers without exposing technical detail.
Why Pentest Machine
- The founder runs tests and owns quality. 20+ years of experience, an international career at six of the world's largest companies and talks at OWASP AppSec USA, Black Hat and several BSides. He takes part in testing and reviews every report before delivery.
- Only senior specialists attack. In-house tooling and AI speed up repetitive work; deciding what to exploit and how to chain it stays human.
- Commercial transparency. Fixed price within 24h, NDA before scoping, retest included and invoicing in the US, EU or Brazil.
- A report for three audiences: the board, developers and auditors.
What you get
- A report with demonstrated findings, business impact and severity.
- Architecture recommendations: authorization before context, least-privilege tools, human confirmation for sensitive actions, usage limits.
- Retest included.
Frequently asked questions
How much does an AI penetration test cost?
It depends on data sources, connected tools and roles. Testing an LLM, RAG or agent application typically costs $24,000–$80,000 in the US (€14,000–€44,000 in Europe). Fixed price within 24 hours.
Do you test systems built on OpenAI, Anthropic or Gemini?
Yes. We focus on the system you built on top of the model — prompts, data, tools and integrations — whatever the provider.
Do we need to share source code?
No. We recommend access to the prompt and tool configuration plus test accounts per role.
Does this help with the EU AI Act?
The report documents robustness and cybersecurity testing, useful evidence for AI risk-management obligations.
Further reading
Other services
Request a quote
Get a proposal within 24h.
Tell us what you need tested. An expert — not a salesperson — replies with scope, timeline and price.
- NDA available before any detail is shared
- Support in EN, PT and ES
- Invoicing in the US, EU or Brazil