Generative AI / Gabor Horvath / 6 min read
Anthropic’s disclosed incidents: Claude models submitted real forms, bypassed access limits and exploited software flaws. Why an AI agent’s permissions have to be limited technically, not just by instructions.
Generative AI / Gabor Horvath / 5 min read
A Yale-led study finds that the number strings inside language models hide a symbolic structure, and that editing it changes the model’s behaviour predictably. Why this matters for understanding and controlling AI.
Generative AI / Gabor Horvath / 3 min read
As a security engineer, I spent the past months learning how generative AI actually works. Here is what I understood, and where it genuinely speeds up my work.
