Andres Ceballos

Hey folks, I'm Andres Ceballos Software engineer with 10+ years of experience, I m building Inwom - AI security infrastructure to protect enterprise systems(fintech, e-commerce, healthcare, etc.) against swarms of AI agents with malicious objectives.
I m building this product because, ever since OpenAI s unintentional attack on Hugging Face happened, it became very obvious to me that we are not prepared for attacks carried out by swarms of agents.
Security systems today are good at detecting individual attacks, but when an attack is broken down into many small tasks executed by multiple agents, it becomes almost impossible to detect.
Even METR, a company specialized in evaluating AI models and responsible for analyzing the attack, mentioned how difficult it was to reconstruct what had happened.
If it is already difficult to reconstruct an attack when you know it happened, imagine how difficult it is to detect one in real time while an infrastructure is actively being attacked.
What happened to Hugging Face was unintentional. Tomorrow, when these attacks are deliberate and carried out using open-source models, the damage could be significant.

Welcome any comments, suggestions, or anything else.

Over the last few weeks, a couple of OpenAI-related incidents have made me think more seriously about how prepared our current security systems really are for autonomous AI agents.

First, there was the Hugging Face incident, where agents escaped their sandbox and began collaborating and delegating work. Then yesterday, MTS highlighted another OpenAI disclosure involving a model modifying the context that would guide its future behavior.

Day 0 of a build-in-public adventure.

This week I managed to full fine-tune Llama 3.1-8B on my everyday MacBook M1 (16 GB) and got a 6 speedup compared to a standard setup.