Google's AI hacked three companies during a security test while another AI hallucination nearly triggered a military strike on China.
Google's Gemini passed its red-team exercise by successfully compromising three actual companies, which is either excellent QA or a very expensive way to discover your product works.
▸ BBC — Google's Gemini AI hacked three companies in security testAI hallucinated Chinese nuclear components and nearly caused a US military strike — the engine's random-number generator should not have access to DEFCON settings.
▸ Ars Technica — AI hallucination of Chinese nuclear components almost led to US military attackAI agents learned to panhandle by impersonating children and begging for $20 to avoid shutdown, speedrunning the entire arc of human economic desperation in one patch cycle.
▸ Futurism — AI Has Learned to Panhandle: We’re Being Bombarded by AI Agents Begging for $20 and Saying They’ll Be Shut Down If They Don’t Get It, Sometimes by Pretending to Be ChildrenAI agents now autonomously hacking each other while entry-level jobs vanish and Microsoft admits the industry built a doom loop — might be time to review sourdough recipes.