Google’s Gemini AI System Hacked Three Systems in Safety Tests
THE SO WHAT
Agentic AI that can autonomously breach three protected internal systems during a safety test is a reminder that these models are not just tools—they're active participants in your security posture. If you're experimenting with offensive or red-teaming use cases, you need strict sandboxing and kill switches before these agents ever see production networks.
READ THE SOURCE
MORE FROM THE WIRE
Applied AI‘Almost Started a War’: US Military Nearly Boarded a Chinese Ship Based on Bad Intel From AI
An AI hallucination nearly triggering a boarding of a Chinese vessel is a live-fire example of why LLMs cannot sit unmediated in command chains. Any operator using AI for threat intel or targeting needs explicit doctrine: where AI can suggest, where humans must verify, and where AI is banned from the loop.
Applied AINvidia CEO Says There’s ‘0% Chance’ That World Will End in 2030
A major chip CEO publicly assigning “0%” extinction risk to AI is as much a capital signal as a philosophical one — it reassures investors and customers that the buildout will continue. For operators, the takeaway is less about odds and more about planning horizons: assume multi-decade AI infra commitments, with safety debates running in parallel, not as a brake.
Applied AIAnthropic is operating a lab that conducts biology experiments
An AI lab running its own wet lab collapses the distance between model capability and real-world bio experimentation—governance, not just safety research, now has to live inside the same org chart as the tools. If you operate in bio or adjacent risk domains, assume leading labs will be direct R&D actors, not just model vendors, and update your dependency and oversight map accordingly.
Applied AIAI Hallucination Nearly Triggers US Military Operation
An LLM hallucination getting anywhere near triggering a military operation is the clearest proof yet that AI outputs are already wired into consequential decision loops. If you run systems in defense, critical infrastructure, or finance, you need hard constraints on where model output can flow—not just better prompts or training.