AI pentesting in Chile and Latam
A classic pentest does not cover a chatbot or an agent with tools. We test models, RAG, guardrails, and MCP servers the way an attacker would, with reproducible evidence and a plan that works for ANCI, Law 21.663, and Law 21.719.
We attack your AI the way a real adversary would, and leave you in order.
A team of experts
Our work shows up inside large companies across Latam.
We do not sell a prompt scanner: we assess with the frameworks customers, auditors, and the regulator already ask for.
- OWASP Top 10 for LLM and agentic applications
- MITRE ATLAS (adversarial techniques against AI)
- NIST AI RMF and ISO/IEC 42001 as a governance map
- Evidence usable in due diligence and under Law 21.663 / ANCI
Typical scope of an LLM pentest or agent red team in production.
- Direct and indirect prompt injection (RAG, tickets, documents)
- Jailbreaks, system-prompt leakage, and data exfiltration
- Tool abuse, function calling, and MCP servers
- Support chatbots, internal copilots, and autonomous agents
What this looks like in practice
AI outside the inventory: what Shadow AI is, why it breaks security and AI governance, and how to start governing it.
We define the category in Chile and Latam: what gets tested, what a web pentest does not cover, and when you should commission one.
Real injection patterns and controls that work in production, not only in the system prompt.
Endpoint, transport, and authentication of MCP connectors so you can inventory tools before an LLM pentest.
Frequently asked questions
What is an LLM pentest?
- It is an offensive test of a system with a language model: chatbot, copilot, or agent. It is not limited to scanning the web app: it evaluates prompt injection, data leakage, guardrail bypass, and tool or MCP abuse, with reproducible evidence.
How is it different from a classic pentest?
- A web pentest covers authentication, APIs, and infrastructure. An LLM pentest covers the AI layer: system instructions, RAG, connected tools, and what a malicious user can achieve by talking. At onenorma they can be hired together.
Does it help with Law 21.663, ANCI, or Law 21.719?
- Yes. We deliver prioritized findings and evidence you can use as a technical control and as proof of diligence before ANCI, cybersecurity framework law 21.663, and the security duties in Law 21.719.
Do you work with companies in Chile and Latam?
- Yes. The service is built for teams in Chile and Latam: reports in Spanish, frameworks local customers ask for (OWASP LLM, MITRE ATLAS, NIST AI RMF), and an initial diagnosis at no cost.
The pentest and the executive report from the onenorma team were very detailed, letting us act quickly and mitigate every finding in a straightforward way. Communication was fast and very clear.
Ready when you are.
Book your strategy session or email us: we reply with a clear scope and the next step.
30 minutes · No commitment · You leave with an action plan for the deals that stall