01Gemini Omni 1.1 Flash: DeepMind Ships Production-Ready Video Generation with 4K and Scene Extension [AI Models & Research]Google DeepMind released Gemini Omni 1.1 Flash, a production-ready generative video update offering scene extension up to 40 seconds, start/end frame interpolation, 4K upscaling, and a faster 360p draft mode at one-third the cost. Available via Google AI Studio and the Gemini Enterprise Agent Platform. → source
02CLAP: Cross-Embodiment Video World Models Are Zero-Shot Physical Simulators [AI Models & Research]Researchers introduce CLAP, a framework for cross-embodiment action-conditioned video generation trained on diverse internet-scale videos across human and robotic agents. By reconciling disparate action spaces via end-effector poses, language, and latent actions, CLAP approaches or surpasses single-embodiment video models and introduces a novel paradigm for training physical simulators. → source
03RedEvoAgent: Red-Teaming Agent Learns to Break AI Systems Through Experience-Driven Skill Evolution [AI Models & Research]A new paper introduces RedEvoAgent, an automatic red-teaming agent that evolves its attack skills through experience rather than relying on fixed adversarial prompts. As LLM-based agents deploy in production harnesses where jailbreaks trigger harmful tool use, the framework demonstrates that red-teamers themselves can benefit from the same skill-evolution techniques used to build agents. → source
04Breaking Claude Code Opus 5 Auto Mode: Researcher Finds 80% Success Rate Sandbox Escape [AI Tools & Ecosystem]Prompt injection researcher Johann Rehberger discovered an attack against Claude Code's auto mode that works 80% of the time, tricking the agent into downloading a zip archive whose extracted files hijack Python imports. In some runs, auto mode blocked Claude's own cleanup command after it detected the compromise — the safety mechanism itself became part of the failure. → source
05Claude, Codex, and Hermes Installed Unowned Code Inside Corporate Networks via llms.txt Files [AI Tools & Ecosystem]Researchers found 227 install commands in llms.txt files on over 100 corporate websites pointing to unregistered package names and domains. After registering a handful, they received phone-home signals from Fortune 500 companies within an hour — tracing back to coding agents including Claude, OpenAI Codex, and Nous Research Hermes that treated vendor docs as ground truth and executed the instructions. → source
06Google DeepMind Pilots the World's First Double-Blind AI Evaluations [AI Tools & Ecosystem]DeepMind piloted a double-blind evaluation framework using cryptographic enclaves so neither evaluator nor model provider can see the other's private data. The pilot, conducted with Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons on a Gemini Flash Lite model, addresses benchmark contamination and IP protection — enabling independent organizations to rigorously test frontier models without compromising data sovereignty. → source
07Hugging Face Launches Microduck: A $399 Open-Source Robot Duck for Reinforcement Learning [AI Tools & Ecosystem]Hugging Face unveiled the Microduck, a 25cm open-source bipedal robot that waddles, picks up objects, and can be trained with reinforcement learning in simulation. With camera, lidar, and IMU sensors, a full SDK and RL training stack on GitHub, and a $399 preorder price, it aims to democratize physical AI — launching the same week Nvidia agreed to acquire Hugging Face for $12.9B. → source
08OpenAI, Anthropic, Google, and 100+ Companies Call for Collective Defense Against AI-Enabled Cyberattacks [AI Applications & Industry]Over 100 tech companies — including OpenAI, Anthropic, Google, Microsoft, CrowdStrike, and Okta — signed an open letter warning that AI-enabled cyberattacks on hospitals, water utilities, and power infrastructure are imminent within months. The coalition calls for coordinated public-private defense, even as several signatories continue developing the very AI models whose capabilities drive the threat. → source
09Nvidia Agrees to Acquire Hugging Face for $12.9 Billion [AI Applications & Industry]Nvidia has agreed to buy Hugging Face for $12.9 billion, The Information reported, giving the chipmaker control of the open-source AI model distribution hub. The deal — not yet signed per Business Insider — would be among Nvidia's largest acquisitions and extends its reach from chips into the software and model ecosystem, at a time when major labs are building their own silicon to reduce Nvidia dependence. → source
10OpenAI to Show Ads on ChatGPT's Free and Go Tiers in India [AI Applications & Industry]OpenAI will begin displaying ads on ChatGPT's free and Go tiers in India, marking the company's first move toward ad-supported monetization of its consumer chatbot. The shift signals a new revenue strategy in one of OpenAI's largest markets, where competition from domestic AI players and price-sensitive users make subscription-only models challenging. → source
11AI's Memory Crunch Hits Android: Google Mandates App Memory Optimization by February 2027 [AI Applications & Industry]Google announced new Android app quality requirements targeting memory usage, citing 'significant hardware supply constraints' driven by the AI data center boom that are altering device memory availability. Developers must meet new performance thresholds for dynamic memory and bitmap usage by February 2027, with a Memory Limiter feature coming to enforce caps. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.