LLM flaw exposes hack risks as geothermal plant gets revived

A research team presenting at a top AI conference argues that large language models cannot be made fully secure against hacks because of a fundamental weakness in how they determine who or what is giving instructions. By exploiting that flaw, the researchers prompted popular LLMs to provide blocked information, including how to synthesize cocaine and how to sabotage a commercial aircraft’s navigation system.

Zanskar has revived the Lightning Dock geothermal power plant in New Mexico after buying the struggling facility in June 2024. The plant had become uneconomical as water from its underground reservoir cooled, but advanced modeling and modern drilling helped the company identify a better well site, drill down thousands of feet, and restore operations to full capacity.

Other technology developments include an FCC ban affecting Roombas and other foreign robots, Google DeepMind’s move away from its AlphaFold team, rising infrastructure pressures around data centers, DoorDash’s drone delivery plans, and growing concern that AI is accelerating global digital inequality. Europe is also testing Project ASGARD, a digital targeting web designed to connect battlefield sensors and shooters through a shared automated network.