Search AI Products and News

OpenAI on Tuesday issued new safety policies aimed at preventing incidents during model testing, strengthening development monitoring and post-training alignment. It is the first major change since the Hugging Face incident; reasons include Astra's cybersecurity capabilities and rapid progress. Measures include pausing frontier reinforcement learning and a 30-minute early alert.....

A self-represented plaintiff in Connecticut, USA, concealed hidden instructions in court documents with white text on a white background, only readable by AI, attempting to influence the case handling. It became the first case of "prompt injection" in the U.S. judicial system. However, the state court did not actually use AI to review the materials, so the method did not work, and the party was penalized for abusing the procedure.

OpenAI engineer Tibo announced that the 1 million Token context window of the GPT-5.6 Sol model in Codex has been expanded from being limited to API keys to being available for ChatGPT accounts. Subscribers can manually activate the million-token window, solving the limitation of being forced to truncate historical records when processing large projects.

DeepSeek-V4-Pro-0813 debuts Day-0 on SiliconFlow, featuring 1M context and low/high/max inference modes, focused on coding, tool calling, and agent workflows, MIT open source. Pricing: $1.32/M input, $3.96/M output, $0.44/M cache hit, with notable cache cost advantage.....

Kingsoft Office Lingxi Pro integrated DeepSeek-V4-Pro on Aug 14, 2026. Users switch manually without setup. Supports 1M token context, max output 384K tokens, benchmarks improved over Preview, differentiated real-world advantages.....

Developer Jeff Johnson exposes Apple App Store rating fraud: TabControl, #12 paid Safari extension in the US, uses suspected AI-generated promo images and a fake 4.9-star rating, exploiting review blind spots to deceive users.....

US frontier AI cybersecurity reviews currently cover only closed models but will almost certainly extend to open-source models in coming months, Wired reports. Frontier refers to models comparable to Anthropic Claude Mythos or OpenAI GPT-5.6. The framework remains voluntary; some officials fear mandatory reviews could stifle AI development.....

Cursor is launching Origin, an automated code review platform, ahead of its original fall timeline. Built by the acquired Graphite team, it aims to break GitHub's single-diff, sequential merge model, addressing review and merge bottlenecks caused by AI agents generating code at scale in parallel. Core modules like Codebase will be added to the web app.....

Anthropic to launch Sonnet 5.5 (codenamed Fennec) next month, featuring 2M token context, fast low-latency reasoning, improved multi-step planning, and enhanced browser/terminal tool use. Performance rivals flagship Claude Fable 5 but retains Sonnet pricing for best value.....

Tencent Hunyuan launches the speech recognition model Hy ASR 3.0 preview, which leverages the language comprehension capabilities of the large language model Hy3, combining high-precision recognition with deep semantic understanding. It has been significantly improved in general recognition, context awareness, multi-scenario robustness, and dialect coverage. It can provide more accurate and coherent transcriptions by combining context, achieving an evolution from word-by-word transcription to semantic understanding.

Tencent Hunyuan launches Hy ASR3.0preview, integrating Hy3 LLM's semantic understanding into speech recognition for context-aware, direct transcription. On open benchmarks, word error rates reach 3.34% for Mandarin, 2.62% for English, and 3.12% for Cantonese, achieving around 3% overall with impressive multi-dimensional performance.....

OpenAI launched the open-source Codex Security CLI tool on July 29. It is a code security scanner that detects and verifies vulnerabilities in repositories, integrates into CI/CD pipelines. Discovered before release, it's an early version; OpenAI will iterate based on developer feedback.....

Black Forest Labs released FLUX3, a unified multimodal model for joint image, video, and audio learning. Built on the Self-Flow self-supervised flow matching framework, it extends FLUX series for multimodal generation and understanding. Supports text-to-video, image-to-video, generating up to 20s videos with native synchronized audio, outperforming predecessors.....

Black Forest Lab releases Flux3, a multimodal model built on Self-Flow architecture, integrating image, video, audio, and motion codecs for unified understanding and generation of physical and digital worlds. It is the first to support native audio generation, producing 20-second synced audio-video clips in one pass. Capabilities include text, image, and video-to-video tasks.....

Black Forest Labs unveils Flux3, a multimodal foundation model using Self-Flow architecture with dedicated encoders/decoders for image, video, audio, and motion. It unifies understanding and generation across physical and digital environments, outperforming Luma and Runway. Flux3 natively generates audio, outputs 20s sync’d video with audio, and supports text/image/video-to-video, keyframe transitions, and multilingual dialogue.....

German startup Black Forest Lab released Flux3, a multimodal foundation model with Self-Flow architecture. It integrates image, video, audio, and action codecs for unified understanding and generation of physical/digital environments. As the first model with native audio generation, Flux3 outputs 20-second synchronized audiovisual clips and supports text/image/video-to-video and keyframe-driven tasks, significantly advancing generative video tech....

Poolside released the fully open-source Laguna S2.1, a 118B-parameter MoE model with 8B active parameters per token and up to 1M context length for long-cycle software engineering. It's freely available on OpenCode and features thinking and no-thinking reasoning modes, sparking developer buzz.....

Ahead of the 2026 World AI Conference, Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open-source model, now the world's largest open-source model, surpassing closed-source rivals in size for the first time. It's not just about scale—it marks a significant open-source breakthrough.....

AI-generated code surge strains open-source review. Godot Foundation bans autonomous AI agents and 'Vibe Coding,' requiring manual coding, accountability, and fixability to ensure project quality.....

Xiaohongshu open-sourced RedKnot inference engine, solving speed and cost challenges in long-text generative AI. Its core innovation breaks token-dimension storage in traditional KV Cache, offering a cost-effective solution for long-context tasks.....