Sovereign AI Lab for Cybersecurity
Independent research. Sovereign intelligence. Advanced cyber capabilities.
We research and engineer AI that organizations and governments can own, operate and evolve themselves.
We are an applied security research firm, focused on developing effective offensive technology to support critical operations.
01 · Where we come from
It started with research.
LESIS was founded by people who wrote their own fuzzers, analyzers and exploits to answer what off-the-shelf scanners could not. When the instrument we needed did not exist, building it was the work.
02 · What changed
Then the instrument changed.
Language models are now the most powerful one security has ever had. They read code at a scale no team can match, and the strongest of them, open only to verified defenders such as LESIS, find flaws that stayed hidden for decades.
Yet almost everyone uses them the same way: rented from a handful of providers, on their terms, with code and incidents sent wherever the model runs. Intelligence you can only rent belongs to whoever owns it, and so does each decision built on it.
03 · What we believe
Cybersecurity requires intelligence that organizations can understand, control and trust.
Understand
Each result carries its evidence, and whatever a model contributes is marked as the model's. A person can always see what was found, how and by whom.
Control
Whoever runs a model decides what it sees, what it keeps and when it changes. That has to be the organization that depends on it, not the vendor selling it.
Trust
Trust comes from proof, not fluency. A model may propose, but nothing counts as a finding until a person has checked it and answers for it.
04 · What sovereignty means
Independence you can inspect.
Five things make it real: the models, the data, the infrastructure, the intellectual property and the way AI capabilities evolve. None of them requires a foundation model built from scratch. Here is where we stand on each.
Models
Model-agnostic by design. LESIS runs closed models, such as Claude and GPT, and open-weight models, such as DeepSeek and GLM. You choose which one handles each task; you stay in control. Specialized models for security work are adapted from open-weight ones and tested on real findings.
Data
Secret values are masked before any model sees them, repositories are connected read-only, and each model call is logged with the model, the task and the cost. Pipelines keep sensitive material where it belongs, so only what the organization approves ever leaves.
Infrastructure
The platform and open-weight models, such as DeepSeek and GLM, run inside the organization's own infrastructure, and nothing leaves it. A full cloud installation, from network to backups, is written as code. Inference runs on infrastructure operated by LESIS, including installations on networks cut off from the internet.
Intellectual property
The platform, its analyzers and our methods are made in house, so what we learn along the way stays with us. Datasets drawn from our research are the raw material our specialized models are made from.
Evolution
Each time an analyst accepts or rejects an AI result, it is logged, giving us a running record of where a model is right and where it is not. Models are evaluated, adapted and redeployed on our clients' schedule, not on someone else's release calendar.
05 · Where we apply it
Only cybersecurity.
A model trained on real findings, exploits and fixes knows things that one trained on the open internet does not. So the lab works on security and nothing else.
Offensive security
Models and tooling that find and chain vulnerabilities, steer fuzzers toward paths others miss and validate exploits in isolated environments.
Defense
Code review, root causes that close several findings at once, and retests that confirm a fix still holds as the code keeps changing.
Cyber intelligence
Findings, assets, incidents and attack paths brought together into a single picture of exposure an organization can act on.
Evaluation
Benchmarks and test sets for security tasks, so any model we use or build is judged on the work it is actually meant to do.
06 · Track record
Work you can verify.
Vulnerability disclosures
CVE-2020-9376 and CVE-2020-9377 in D-Link, CVE-2020-28043 in MISP, and reports credited by Oracle and Apple.
Competitions
First place in the BSides São Paulo CTF, a Hackaflag qualifier win, and the LESIS-backed GRIS/UFRJ team first among 18 at SBSeg 2026.
Open source
Security Gate, Bunkai and Kasumi, tools the team builds and maintains in the open, for anyone to use and audit.
Talks
Presented at BxSec, Roadsec, BSides Rio de Janeiro and XibéSec, and shared with the wider Brazilian security community.
Your infrastructure. Your intelligence. Your sovereignty.
Modern AI, without handing your most sensitive systems to someone else.
Sovereign Intelligence. Secure Futures.
