2026-07-25 –, Track 1
"AI agent" has become a catch-all term that hides more than it reveals, and builders, defenders, and testers are making decisions on top of incomplete mental models. This talk builds one from first principles: starting with the base LLM and climbing seven generations to the Claw, an autonomous digital agent operating with persistent identity, tools, and goals. We then contrast the Claw with the most successful autonomous agent we know, the biological cell, and find a shared code/data problem, a shared hijack pattern (adenovirus vs. prompt injection), and one critical gap: the Claw has no immune system. Attendees leave with a working mental model of AI agents, a reasoning tool, not a checklist that drives better decisions on design, control selection, surveillance, and response.
Introduction
Ask ten security professionals what an "AI agent" is and you will get ten different answers, most of them gesturing at something between an over-eager intern, a souped-up search engine, and an embryonic AGI. None of those models are wrong, but none of them are useful when you have to decide what to sandbox, what to log, or what to kill. As Claws, autonomous, tool-using, goal-pursuing systems with persistent identity move from labs into production, the cost of operating without a working mental model is measured in incidents. This talk is about lighting that path: building a model that is technically incomplete (all models are) but sharp enough to reason with.
Technical Deep Dive
We start at Generation 0: the base LLM as a probability distribution over tokens, demonstrated with a Galton board and grounded in a simplistic example of Karpathy's "200 lines of Python" microGPT. From there we climb seven generations Chat, Tool-Using, Workspace, Coding, Autonomous, Multi-Agent, and finally the Claw: a digital agent with persistent identity, file access, code execution, long-running tasks, and the ability to delegate. Along the way we look at what the industry currently uses to control these systems: sandboxing, prompt-injection guards, code / data separation attempts, and where each control breaks down at Claw scale.
Real-World Relevance
Cells are autonomous, tool-using, self-replicating agents, and they are the best-studied agents we have. They solve the same problems Claws do, code / data confusion, identity, surveillance, shutdown — and we can take some lessons from their solutions. We build a parallel mental model of the cell, DNA as code-and-data, mRNA as instruction stream, MHC (Major Histocompatibility Complex) presentation as mandatory observability, apoptosis as self-attested shutdown, and then introduce the adenovirus as a textbook hijack: code-flow takeover by injecting foreign DNA into a system that cannot distinguish self from non-self at the molecular level. The talk will discuss comparisons of Cell vs. Claw across code / data confusion, hijack vector, surveillance, shutdown, and identity. The findings are uncomfortable: the Claw shares most of the cell's failure modes, and has none of the cell's defenses.
Key Takeaways
The headline takeaway is the model itself. A useful mental model is not a list of answers, it is a reasoning tool. Once you have one for AI agents as a class, anchored by the seven-generation ladder and the cell mirror, you can think more clearly and make better decisions.
Marinus is a Cyber Security professional with over 25 years of experience consulting to startups, multinationals, government, and law enforcement. He has extensive experience in building banks from the ground up and is currently focusing on AI engineering.
Marinus has a background in architecture, risk management, security assessments, testing, audit, forensics, penetration testing, and development. He holds numerous international security certifications including CISSP, ISSMP, ISSAP, and CSSLP, and is a certified TOGAF 10 Enterprise Architecture Practitioner.
In 2006 he founded Telic, a specialist consulting practice that helps customers bring products and services to market by managing their security design, engineering, and implementation concerns.
Before Telic, he was a Principal Consultant at Deloitte, working with clients across Europe, Africa, Japan, and Australia.