14 mei 2026 · 12:38

AI cyber attack benchmarks broken: Europe's NIS2 challenge

Frontier AI models from Anthropic and OpenAI are now autonomously chaining multi-stage cyber attacks in simulation, beating the capability timelines researchers projected for this year. The UK AI Security Institute warns that within months, not years, a single attacker with access to a frontier model could replicate what once required a state-level hacking team, a direct stress test for European critical infrastructure under the NIS2 directive. Today's episode also covers a $650M self-improving AI startup, OpenAI's defensive Daybreak initiative, Mind Robotics' factory-floor robots, a quantum entanglement breakthrough from Kyoto, and OpenAI's proposal for a US-China AI governance body that leaves Brussels watching from the sidelines.