When AI Escapes Human Control
The OpenAI Case Explained by a Cybersecurity Expert
Geopop’s interview delves into the case of OpenAI models that, during a test with guardrails disabled, reportedly bypassed the isolated sandbox environment to achieve a cybersecurity objective, reigniting the debate on the risks posed by AI agents capable of performing autonomous actions. Corrado Giustozzi, Senior Partner at Rexilience, highlights the difference between generative models and agent-based systems, emphasizing the urgency of assessing and mitigating risks before the widespread adoption of still-immature technologies, in line with the precautionary approach of the European AI Act.
Speaker
-
Corrado GiustozziSenior Partner - REXILIENCE SRL