Security Research · Web · Mobile · AI

Finding and documenting how systems break.

I'm Liliane Zukerman, a security researcher with a background in offensive security and malware analysis. I work on the vulnerabilities that matter in web applications, mobile platforms, and AI-based systems, and I write up what I find.

Read the research →
01 / research

Vulnerability research & disclosure

Reproducible analysis of real vulnerabilities: root cause, proof of concept, and practical guidance for detection and mitigation.

Israel's 2026 election: what the ballot-box data show
Independent audit of the 26th Knesset election with the Central Elections Committee's results for all 12,929 ballot boxes, against 2022: digit tests, turnout, same-site comparisons, the double-envelope count and the Bader-Ofer seat allocation. The count checks out; eleven transcription slips flagged on the preliminary file were all corrected in the final one. A 100% paper election audited with the same method as Brazil's 100% electronic one.
Scripts and data: github.com/liliane-zukerman/israel-election-audit-2026
Eleição presidencial de 2026: o que os dados públicos do TSE mostram, e o que eles não conseguem mostrar
Auditoria independente do 2º turno com dados públicos do TSE: a soma dos 497.897 boletins de urna bate voto a voto com o resultado oficial, os casos que circularam nas redes se explicam pelos próprios dados, e 1,7 milhão de votos para presidente foram anulados por engano. Testes forenses, logs, RDV e assinaturas de software, com scripts e dados abertos.
Scripts e dados: github.com/liliane-zukerman/auditoria-eleicoes-2026
CVE-2026-59726 (RufRoot): Unauthenticated RCE in the Ruflo MCP Bridge
Independent analysis of unauthenticated remote code execution in the Ruflo MCP Bridge via missing authentication on the JSON-RPC endpoint, enabling arbitrary command execution and LLM API key theft. Ruflo < 3.16.3. CVSS 9.8 CRITICAL · CWE-306.
CVE-2026-88028 / CVE-2026-88027: Query-Operator Injection in Laravel MongoDB Relation Identifiers
Root cause analysis and proof of concept for unauthorized document disclosure via query-operator injection in mongodb/laravel-mongodb < 5.11.0. CVSS 7.1 HIGH · CWE-943.
02 / ai security

AI & LLM security

Offensive research on AI-based systems: prompt injection, jailbreaks, and the infrastructure around LLMs and MCP. Testing how model guardrails and the applications built on them actually hold up.

LLM01: Prompt Injection in DVAI, From a Toy Simulator to a Real Model
Reverse-engineering the three prompt-injection challenges in Damn Vulnerable AI against both a scripted detector and a live model. Why keyword guardrails fail, where instruction defenses break, and the inverse relation between how famous a jailbreak is and how well it works.
Also available in: Português
Prompt Injection Testing Checklist
A practical, repeatable checklist for testing an LLM application against prompt injection, with ready-to-use payloads tagged by whether they beat a real model, a keyword filter, or both.
03 / about

About

I'm passionate about technology and security. Over close to two decades in the field, I've moved from infrastructure and network administration into penetration testing, offensive security, and malware analysis and reverse engineering.

Today my focus is research across web, mobile, and AI systems. My academic background started in physics and moved into computer science. I'm fluent in English and Portuguese, and I speak basic Hebrew.

How I work. The research on this site is done with AI assistants (Claude, among others) as a working tool: drafting, writing and reviewing code, running analyses and checking numbers. Every result is reproduced by me from the data, every script is published so that anyone can rerun it, and the conclusions and mistakes are mine.

Certifications

Education

04 / contact

Get in touch

Open to security research collaboration, consulting, and conversations about interesting vulnerabilities. The best way to reach me: