Pentest de LLM e Features de IA
Auditoria de feature com LLM — prompt injection, prompt leak, jailbreak, custo aberto, drain de chave, RAG poisoning, hallucination com impacto.
Why now
The real pain
Empresa adicionou chatbot com Claude/GPT em produção em 2 semanas. Não testou prompt injection, system prompt leak, drain de chave, abuse via PII. RAG conectado a banco interno vaza dado privilegiado. OWASP LLM Top 10 é o novo OWASP API.
Applicable frameworks
/attack-surface
Pentest LLM e IA
Every engagement is designed for your environment. The points below are part of our standard playbook for this sector — final scope is adapted to your stack and contract.
Prompt injection (LLM01)
Direct e indirect injection, system prompt override, jailbreak via roleplay/encoding.
Prompt leak (LLM07)
Extração do system prompt, vazamento de regra de negócio, leak de exemplo few-shot.
Sensitive Information Disclosure (LLM06)
Vazamento de PII, dado de outro tenant, dado de treinamento via memorização.
Supply chain & model poisoning (LLM03/LLM05)
Modelo pinned, fine-tune malicioso, RAG corpus contaminado.
Excessive agency (LLM08)
Agente com tool de write privilegiado, sem confirmação humana, ação destrutiva.
Custo, abuse e DoS (LLM10)
Drain de chave OpenAI/Anthropic, abuse de cota gratuita, regex DoS via prompt.
/methodology
Genuinely manual pentest
Automated scanners find what's documented. Real attackers find what isn't. 90% of the work is manual — performed by specialists holding OSCP, CISSP, CRTO and GPEN.
01 · Reconnaissance
Target mapping, OSINT, footprint, sector-specific threat modeling.
02 · Discovery
Deep enumeration, complementary scanning, manual exposure identification.
03 · Exploitation
Manual validation with controlled PoC, finding chaining, escalation.
04 · Report
Executive + technical, step-by-step replication, mapped to applicable regulation.
/why-trust
Who has trusted our work
Auditoria de features GenAI em SaaS B2B, fintech (KYC com vision LLM) e healthtech.
Technical assessment recognized in highly regulated, mission-critical environments — the pentest that finds what nobody had found before.
Douglas Lopes
Founder · CEO · intrus.io
/faq
FAQ — Pentest LLM e IA
Vocês cobrem agentes autônomos?
Sim. Agente com tools de write (banco, email, transação) é categoria mais perigosa. Auditamos confirmação humana, sandboxing, isolamento de tool.
Faixa de preço?
Feature LLM única (chatbot/assistant): R$ 12-30k. Plataforma com múltiplos LLMs e RAG: R$ 30-90k. Agente autônomo com tools: R$ 50-150k.
/contact
Cotar pentest LLM
Schedule a confidential meeting. Within 48h we'll send a proposal with scope, timeline and pricing.