AI Architect; AI Security

About the project (description, duration, stage)

Hands-on AI-for-Security engagement with a regulated iGaming / online-gaming group
. The client’s security team is genuinely advanced: they already run an AI-driven offensive-security capability — continuous external-perimeter scanning feeding an LLM agent that plans exploitation, sources and validates exploits, and executes them in sandboxed environments — plus a runtime anomaly-detection layer watching for intrusion and privilege-escalation patterns across their products. They built this themselves and have explicitly asked us to challenge and improve it, not just rubber-stamp it
.

This is not a generalist AI project
. Neurons Lab brings the AI-architecture and engagement depth; what’s missing is the offensive-security domain lead who can sit across the table from a hands-on CISO team as a peer, pressure-test their pipeline, and own the methodology.
You are that expert. The early work is concrete and consultative: understand what they’ve built, find where it’s wrong or expensive, and propose a better way.

Stage : pre-engagement / discovery (the immediate next step is a joint technical session with the client’s CISO / security engineers).
Duration : discovery → advisory / PoC, with strong extension probability as the security program scales across the group.

Reporting :
Neurons Lab CTO / engagement lead (@Alex Honchar); partners with the Neurons Lab AI Architect on the account. You are the security domain owner for this track.

• Join joint working sessions with the client’s hands-on security engineers;
  challenge and harden their AI-driven offensive pipeline end-to-end (recon → verification → AI-planned exploitation → sandboxed execution).

• Design and refine the exploitation agent : how the LLM plans attack paths, selects and validates exploits, and orchestrates parallel sandboxes safely and reproducibly.

• Optimise cost-per-finding of the existing exploitation pipeline: benchmark local / sovereign open models (Kimi, GPT-OSS, Mini Max, Deep Seek) against frontier models for the recon, exploitation and analysis loops; quantify accuracy / latency / cost trade-offs and recommend hardware sizing.

• Shape the runtime anomaly-detection layer : define which intrusion / privilege-escalation precursor patterns are worth collecting (signal over raw-log volume), and design the missing pieces —
  automated response (kill a malicious process / disable an account on detection) and triage routing by criticality.

• Stand up a quick-win PoC to anchor the engagement — e.g. an automated dependency / PR vulnerability-scanning pass, or a head-to-head local-vs-frontier benchmark of the exploitation agent.

• Turn findings into a defensible technical proposal and roadmap ; present methodology and trade-offs to a technical CISO / CTO audience.

• Keep all sensitive work build-time and in-perimeter — no pushing intellectual property, configs, or recon-enabling data to external model providers; respect regulated-gaming certification constraints (no uncertified AI in runtime-critical paths).

• Hands-on offensive security : vulnerability research, exploit development and chaining, web + network penetration testing; fluent with Nmap, Nuclei, Katana, Acunetix, Metasploit, Burp Suite and Kali tooling.

• Building and operating LLM agents for security work — agentic tool-use, sandbox orchestration, prompt / flow design for recon and exploitation, guardrails for autonomous exploitation.

• Local / self-hosted open models : running and tuning open weights (Kimi, GPT-OSS, Mini Max, Deep Seek) on rented or private GPU; quantization, throughput and the agentic-performance trade-offs that matter for security automation.

• Exploit & threat intelligence :…