Descriptions:
The AI Search channel’s weekly roundup covers one of the most eventful weeks in recent AI news, led by Anthropic’s announcement of Claude Mythos Preview — a model the company describes as having surpassed nearly all human experts at finding and exploiting software vulnerabilities. According to Anthropic’s official blog, Mythos has identified thousands of high-severity flaws across Windows, macOS, iOS, Android, Chrome, Safari, and Firefox, including a 27-year-old vulnerability in OpenBSD and a 16-year-old flaw in FFmpeg. Rather than a public release, Anthropic launched Project Glasswing — a defensive alliance including Google, Nvidia, Microsoft, Apple, and AWS — providing defenders early access to patch systems before attackers can leverage similar capabilities. Benchmark comparisons show Mythos achieving a 14% improvement over the previous state-of-the-art on SWEbench Pro and a 13% gain on SWEbench Verified.
Other major stories include Meta releasing Muse Spark after reportedly rebuilding its AI research team from scratch following disappointing Llama 4 results; Alibaba’s new video generation model Happy Horse — led by the former chief architect of Kling — which appears to outperform most existing video models; and DeepSeek teasing a version 4 model. The video also covers ZAI’s open-source GLM 5.1, described as the top available open-weight model at time of publication.
On the interactive media side, Overworld’s Waypoint 1.5 is highlighted as a real-time world-generation model capable of running at 720p and 60 frames per second on consumer Nvidia GPUs, with an open-source release available immediately.
📺 Source: AI Search · Published April 12, 2026
🏷️ Format: News Analysis
