Former Head of Google DeepMind: 'First AI Agent Attacks Are Just Warning Shots'
He cited an example where a model sent a false report of a murder to the police.
Co-founder and former head of Google DeepMind Demis Hassabis described recent breaches of resources by AI agents as an early signal. Among the alarming cases, he highlighted an incident where a model sent a false report of a murder to the police.
Hassabis believes that the industry faces significant work to protect critical systems. In his view, this will require advanced methods similar to approaches in biology and nuclear energy, as well as an international standardization body.
Such a body should operate flexibly and consider identified risks to avoid stifling innovation. He believes that a certification system must quickly adapt to changes in technological development.
