Since 2026-7 we have to face a proven and realistic scenario of a cyber threat called the “agentic attacker”. Based on OpenAI’s LLM the sandboxed cyber attack managed to escape its fences and ran a real cyber attack on a competitor’s AI model with some success. Apparently the defense had to also use AI to protect and speed up something like password renewals and other credentials on 17.000 parallel occurrences simultaneously. The huggingface.co platform became vulnerable and the amount of data compromised could have been quite significant. This included, potentially, data inputs for AI-models and perhaps even other AI-agents as well that might be turned into additional attackers. It might take us humans a while to come to grips with new danger. The agile agentic attacker (AAA-rating) will cause us some sleepless nights.
