Cryptology Session
Nurdaulet Kaldybek — Multi-Agent Vulnerability Discovery: Orchestrating LLM Agents as a Team
Abstract:
Vulnerability discovery across a large scope scales poorly: a single researcher is limited by time, while a single general-purpose script or LLM agent is limited in depth. This talk explores an approach where the task is distributed among specialized AI agents—each with a distinct role (recon, analysis, hypothesis testing) and specific data—coordinated via a local GitLab instance: issues serve as the task queue, CI/CD pipelines act as the execution engine, and self-hosting ensures that sensitive data never leaves the infrastructure.
The presentation covers the system architecture and the lifecycle of a finding—from initial reconnaissance to a confirmed result—along with a candid assessment of what works and what doesn't, including strategies for handling false positives and agent hallucinations. This talk is designed for offensive security professionals looking to integrate AI agents into their workflows without turning them into generators of impressive-looking but entirely fabricated reports.
Zhandos Boranbay, CEO Qamal Cyber Education — Modern Approaches to Teaching Cryptography Using EdTech.
Daniel Bohannon (Palo Alto, USA) — MaLDAPtive: LDAP Obfuscation, Deobfuscation & Detection.
Zhasulan Zhussupov — Malware delivery via signal processing.
Mukhamzharov Rauan — Optimization of DES symmetric encryption on the NVIDIA CUDA architecture using bitslice vector technology.
Abstract: This paper examines the application of bitslice technology to the hardware-oriented DES (Data Encryption Standard) cipher using GPU-based computing architectures. It analyzes the inefficiencies associated with the classical implementation of bit permutations and table-based substitutions on general-purpose processors. A mathematical approach is demonstrated for representing zero-cost permutations and for the logic synthesis of substitution boxes (S-boxes). A detailed analysis is presented of the system's software implementation in CUDA C++, which combines bit-level vector parallelism with the massively parallel SIMT architecture.
Jubatkanov K. — Learnable Polynomial Activations for Encrypted Inference.
