Attacking Instant Messaging Applications in the LLM Era
4800€ | 12th to the 15th of October 2026
The security research workflow for IM applications has changed. Large language models can now read and classify thousands of decompiled functions in minutes, build protocol clients from documentation, and map attack surfaces across entire applications. But they also hallucinate about code paths that don't exist and confidently recommend exploit strategies against unreachable targets. Knowing where to trust them and where to verify is the new core skill.
This four-day course teaches a methodology for IM vulnerability research and exploitation that generalizes across targets. We use Telegram and WhatsApp as case studies because together they cover the two main scenarios a researcher faces: an open-source application where the challenge is scale, and a closed-source, heavily obfuscated application where the challenge is understanding. Students build an LLM-powered vulnerability research pipeline in Python against both the OpenAI API and the Claude Agent SDK, then turn it against Telegram's animated sticker processing to independently rediscover a real, recently-disclosed vulnerability class.
The pipeline is model-agnostic by design. Inference runs against multiple backends: a dedicated GPU server hosting leading open-weight models (currently GLM-5.1 and Kimi K2.5, subject to change as the field moves), and commercial APIs including Claude via the Agent SDK. Students use multiple models throughout the course. The course treats model selection as a first-class research decision: which model to use for which task, when to switch, and why.
Objectives of the training
IM Architecture Overview
Understanding the Attack Flow
Building Research Infrastructure
Applied Fuzzing and Symbolic Execution
Advanced Static & Dynamic Reverse Engineering Techniques
Find & Exploit Userland Vulnerabilities in Real Targets
The trainer
Who will run this training?
Nitay
Artenstein
Nitay Artenstein is a senior security researcher and the leader of an international research group.
He has been a speaker at various security conferences, including Black Hat and Recon, and has conducted training sessions in Linux kernel exploitation and baseband research.
He suffers from a severe addiction to IDA Pro (at least until he gets used to Ghidra’s GUI), and generally gets a kick out of digging around where he’s not supposed to.
Syllabus
What will we do?
Agenda
Day 1: Reverse Engineering and LLM-Assisted Analysis
- Instant Messaging Attack Surface Overview
- Environment Setup: Self-Hosted Inference and API Configuration
- JEB on WhatsApp: Navigating Obfuscated Closed-Source Code
- The LLM RE Loop: Snapshot, Analyze, Rename, Review, Apply
- Manual vs. LLM Rename Comparison on WhatsApp Media Pipeline
- Building a Naming Review Agent: OpenAI API and Claude Agent SDK
Day 2: Dynamic Analysis and Protocol Dissection
- Advanced Frida Scripting
- SSL Unpinning
- The WebP Trap: Why You Trace Before You Build
- Mapping WhatsApp’s Sticker Decode Path
- 0-Click Path Enumeration: From File Download to System Service (DNG, Paragon PDF)
- Telegram MTProto Sniffer
- Cross-Target Methodology: Open Source vs. Closed Source
Day 3: LLM-Powered Vulnerability Research
- Pipeline Architecture: From 800 Functions to 5 Candidates
- Telegram MTProto Client
- LLM Pipeline: Extract, Pre-Filter, Classify, Verify (OpenAI API + Claude Agent SDK)
- Writing Classification and Exploitability Agents
- Cross-Target Pipeline Execution on Native Libraries
- rlottie Internals, Android Security Model
Day 4: Exploiting a Live Vulnerability
- Running Student Pipelines Against rlottie
- Rediscovery of a ZDI-Disclosed Bug Class
- Malicious Animated Sticker Construction
- 0-Click or Not? Server-Side Mitigations and Their Limits
- Bypass Research: Secret Chats, Custom Emoji, Side-Loading
- Debrief: Vulnerability vs. Exploitability
Prerequisites
- Python or equivalent scripting
- Reverse engineering experience
- Familiarity with Android application architecture
- Claude Max subscription (active)
- Anthropic API key with $200 prepaid credits
Hardware and Software
- Unix-based laptop, 30GB free disk space
- Rooted Android device (provided)
- Dedicated GPU inference server (provided)
- JEB Decompiler
- IDA Pro with Hex-Rays, or Ghidra
- adb, python3, frida-tools
- Administrator/root access
Other trainings
What else might interest you?
Modern Malware OPSEC & Anti-Reverse Techniques Implementation and Reversing
Silvio La Porta & Antonio Villani