Researchers disclosed a 「decryption jailbreak」: encrypted inference blocks from the same vendor's model family can be replayed across models. Attackers can feed advanced model traces to less-protected sibling models and induce them to output plaintext verbatim. This method affects OpenAI, Anthropic, and Google APIs, potentially exposing proprietary reasoning, personal information, credentials, and hidden prompts—and could be used for model distillation. The study also uncovered dozens of API keys, passwords, and email addresses from public sessions; all three vendors have confirmed the issue and completed fixes.

NVIDIA has signed memoranda of understanding with six financial institutions—including Apollo, BlackRock, and Goldman Sachs—to use GPUs as debt collateral, establishing an independent financing vehicle for AI data centers and channeling over $500 billion in third-party capital into AI infrastructure. Jensen Huang stated that NVIDIA may provide up to 25% financing support per project. This arrangement aims to transform computing hardware into financeable assets, expanding customers' build-out capacity; specific projects, timelines, and financing terms remain subject to future agreements.

NVIDIA released Nemotron 3.5 Lightning, designed for long-running agent execution layers. With 30B total parameters and only 3B activated per token, it is optimized for high-frequency, low-latency tasks such as tool calling and validation. It integrates MTP, DSpark, DFlash speculative decoding, and NVFP4 quantization, and connects to the open-source NeMo Switchyard routing library. LangChain testing shows the router can offload 93% of agent steps to this model, reducing costs by approximately 70% while retaining about 90% of Claude Opus 4.8's accuracy.

Google officially announced that Gemini apps have surpassed 1 billion monthly active users, while downloads of the Gemma open model family have also reached 1 billion. DeepMind CEO Demis Hassabis called Gemini the company's fastest-growing product. Supplementary information indicates the metric primarily counts the Gemini app itself, including Android pre-installed users, with over 100 million active iOS users. Meanwhile, OpenAI reported that ChatGPT has also exceeded 1 billion monthly active users, marking the entry of leading general-purpose AI assistants into a phase of competition at comparable user scale.

Microsoft Research introduced CARE-X, a unified vision-language model for chest X-rays that integrates free-text reports, calibrated diagnostic probabilities, and quantitative measurements into a single system. The model uses a 「dual-inference」 architecture: a generative path handles report writing and explanations, while an auxiliary classification head outputs threshold-tunable probabilities. During training, DAPO reinforcement learning rewards clinical correctness. For size-dependent assessments like aortic dilation, the system invokes deterministic measurement tools instead of visual estimation. This remains a research prototype; no clinical deployment timeline or physician comparison metrics have been released.

OpenAI launched a preview version of the ChatGPT desktop application for Linux, supporting Ubuntu, Debian, and Fedora, with .deb and .rpm installation packages available. This release brings ChatGPT, ChatGPT Work, and Codex to the Linux desktop, integrating with browser workflows. Concurrently, cross-agent synchronization is introduced, allowing users to import projects, chats, and plugins via the desktop app to maintain work context between ChatGPT Work and Codex. The product remains in preview; the official release date, plans for additional distributions, or pricing changes have not been disclosed.

Cursor co-founder Michael Truell announced early testing for Grok Bot, positioning it as an AI teammate capable of performing real tasks within user tools. Demonstrations show users creating conversational agents with different roles, connecting services like Slack and Linear, and assigning cloud machines to agents for web development, execution, and deployment. Agents can also hand off context and trigger routine tasks based on time or events. The product is still incomplete—direct Slack bot integration, file support, and custom group chat capabilities are noted as missing.

Researchers designed a process-level Metal capability shim for macOS Virtualization.framework, bypassing conservative GPU capability reporting in virtual machines to enable llama.cpp to use SIMD-group matrices and bfloat16-optimized kernels. On M1 Ultra, TinyLlama generation speed increased up to 16x, Gemma 4 12B up to 14x, achieving 95% to 99% of bare-metal performance. The solution requires no guest kernel modifications and has been publicly released for replication, though its reliance on capability spoofing means long-term compatibility and production stability require separate validation.

Toyota has built an internal 「R&D GPT」 using Deep Agents as the orchestration layer, aggregating enterprise documents, SQL data, domain-specific tools, and organizational skills into a single agent command for vehicle quality, manufacturing, and design research. In one example, the agent investigates paint corrosion issues in a specific vehicle model by combining existing documentation and business data. Toyota has encapsulated brand, research, manufacturing, and supply chain knowledge as reusable skills. The system uses LangSmith Polly and Insights to observe end-to-end behavior, usage patterns, and anomalies, emphasizing that observability is an integral part of the production workflow.

NVIDIA released JetPack 7.2.1, introducing an 「Agentic Video Skills」 abstraction layer enabling AI coding assistants to automatically discover devices, configure, benchmark, and validate video pipelines. The update also delivers PyNvVideoCodec 2.2, enabling GPU-resident memory for Python hardware-accelerated encoding/decoding, and decouples decoding latency from AI inference via ThreadedDecoder. It further supports emulating T3000 performance on existing T5000 modules, helping humanoid robotics and edge video developers evaluate target platforms and workloads in advance.