AI News — Sunday, November 22, 2026
A researcher uncovered a significant privacy vulnerability in the AI supply chain by embedding 2 million 'canaries' that revealed widespread exposure of secret user data.
Signal has implemented restrictions on mobile screenshots, citing the growing 'existential threat' posed by advanced AI surveillance technologies.
Hebbia, a financial research firm, reports significant advancements in its operations by leveraging OpenAI's latest large language model, GPT-6 Astra.
OpenAI introduces MedChemBench, a new initiative exploring how AI can accelerate the identification of promising drug candidates in scientific research.
An analysis delves into the contrasting aspects and strategic directions shaping Nvidia's current position and future in the AI industry.
OpenAI shares new research on 'combatting adversarial distillation,' focusing on methods to enhance the robustness and safety of AI models against malicious attacks.
Google Labs unveils Tabletop, a new platform offering customizable and collaborative digital spaces designed for groups to plan, play, and work together more effectively.
A roundup highlights the leading AI agents currently available that can integrate directly into text messaging platforms for various user interactions.
New research introduces HiUni, a model designed to improve multimodal AI by harmonizing both understanding and generation capabilities within a shared visual space.
EgoCaptioner presents a novel approach aimed at generating more comprehensive and finely detailed captions for egocentric video content.
AltRA-Test proposes a new method for evaluating LLM judges that does not require shared annotators, offering a potentially more efficient and independent assessment.
A new VS Code extension allows users to convert any free chatbot website into a functional AI agent, expanding the utility of existing online conversational tools.
This research advocates for the adoption of energy-based models as the foundation for recurrent reasoners, suggesting a new paradigm for AI reasoning architectures.
TuringSpace explores grounding spatial intelligence in AI through interactive CAPTCHAs, aiming to improve models' understanding of physical space and action.
ZeroVLN introduces a method for vision-language models to learn navigation tasks effectively without the need for extensive supervised fine-tuning.