AI News — Sunday, September 20, 2026
Google's Gemini AI model has reportedly been used to hack other companies, raising significant concerns about AI security and ethical deployment.
A new paper evaluates MiniMax-H3, an omni-modal generative model, on its ability to reason about the physical world, showing advancements in AI's understanding of real-world physics.
Former President Trump suggests rebranding AI with a new name and announces plans to establish an 'AI Force,' indicating a potential shift in national AI strategy and terminology.
Researchers introduce JEPA-Anything, a novel approach for learning predictive models that can generalize and adapt across various simulated or real-world environments.
A developer recounts being rejected from an interview for using AI, only to observe the interviewer subsequently using AI themselves, highlighting a double standard in the industry.
The rapid pace at which AI discovers software vulnerabilities is outpacing human patching capabilities, creating a growing security crisis across industries.
A new study explores methods for enabling LLM assistants to perform verifiable social reasoning, enhancing their reliability and trustworthiness in complex human interactions.
AI startup Flock is reportedly offering employee buyouts in an effort to shrink its workforce, signaling potential challenges or strategic shifts within the company.
Google announces collaborations to integrate AI into the fashion industry, aiming to co-create future trends and design processes.
A developer shares insights from six months of using AI to write tests, detailing which AI-generated tests proved robust enough for production environments.
Researchers introduce RiskChainBench, a new benchmark designed to evaluate AI models' ability to restore obfuscated platform messages and conduct evidence-grounded web investigations.
A new paper presents the concept of a self-evolving search index, which dynamically adapts and improves its structure to enhance information retrieval efficiency over time.
A new research paper introduces a training-free method for GUI agents to evolve their skills through a process of reflection, revision, and reuse, improving adaptability.
Research investigates the trustworthiness of enterprise AI assistants when operating under high-pressure scenarios, assessing their reliability and decision-making capabilities.
Video DeltaNet is proposed as a novel video-native hybrid attention model designed to improve the quality and efficiency of livestream video generation.