AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
Anthropic's Mythos Cybersecurity Bomb & GLM-5.1's Long-Horizon Breakthrough
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's April 8th, 2026. Anthropic just made what might be the most significant AI security announcement of the year. They've launched Project Glasswing, a new cybersecurity initiative built around a previously unreleased frontier model called Claude Mythos Preview. And the capabilities they're describing are genuinely startling. Mythos has already found thousands of zero-day vulnerabilities flaws that were previously unknown to the software developers. We're talking critical bugs in every major operating system and every major web browser. One was a two-seven-year old vulnerability in OpenBSD, which is considered one of the most security-hardened operating systems in existence. Another was a 1-6-year-old bug in FFMPEG, the video encoding library that's used by countless applications. That particular flaw had survived 5 million automated test hits without being caught. On Cybersecurity Benchmarks, Mythos scored 83.1% on CyberGym compared to 66.6% for Opus 4.6. On SuiteBench Pro, it hit 77.8% versus 53.4% for Opus. This is a big jump. Here's the thing though, Anthropic isn't releasing this model publicly. They're limiting access to 12 partner organizations, including Amazon, Apple, Google, Microsoft, Nvidia, Cisco, CrowdStrike, and the Linux Foundation. Another 40 organizations will also get access. The partners will use Mythos specifically for defensive security work finding and fixing vulnerabilities before bad actors can exploit them. Anthropic is putting 100 million US dollars in usage credits behind this, plus 4 million dollars in direct donations to open source security organizations. The framing here is interesting. Anthropic is essentially saying AI models have reached a point where they can outperform all but the most skilled humans at finding and exploiting software vulnerabilities. That capability is going to proliferate whether we like it or not. So we either put it in the hands of defenders first or we wait for attackers to catch up. Project Glasswing is the defensive play. Staying with new model releases, Chinese AI company Z.ai just dropped GLM5.1, their next flagship model designed specifically for long horizon agentic tasks. And this one is open source under the MIT license. The key innovation here is sustained optimization over long sessions. Most models plateau quickly, they apply familiar techniques, make some initial progress, then stall out. Giving them more time doesn't help GLM5, one is built differently. In one demonstration, it optimized a vector database over 600 iterations with more than 6,000 tool calls, eventually reaching 21,500 queries per second, roughly six times what the best models achieve in a single session. On SW Bench Pro GLM5, one hit 58.4%, which is actually state-of-the-art slightly ahead of GPT-5, 4, Opus 4.6, and Gemini Trace Virgila Um Pro. The model weights are available on Hugging Face and Model Scope, and it's compatible with clawed code and open claw. For companies hesitant to use Chinese models, this is worth knowing about. Alright, a few more things worth knowing about today. Google open sourced something called Scion, an experimental orchestration testbed for multi-agent systems. Think of it as a hypervisor for AI agents. It lets you run multiple specialized agents simultaneously in isolated containers, each with their own credentials and workspaces. The idea is that instead of trying to constrain what an agent can do through rules and prompts, you let the agent work freely but enforce boundaries at the infrastructure level. It supports Gemini, Claude Code, Open Code, and Codex. If you're building multi-agent systems, this looks like a useful piece of infrastructure. Intel is joining Elon Musk's Terrafab project. This is the effort by SpaceX and Tesla to build a new semiconductor factory in Texas aimed at producing one terawatt of compute per year for AI and robotics. Intel's involvement makes a lot of sense. They've been searching for anchor customers for their foundry business and now they have two. Building a chip fab from scratch typically costs more than 20 billion US dollars and takes years. Intel's stock was up more than 3% on the news. And finally, a small US startup called RC just released Trinity Large Thinking, a 400 billion parameter open source model built by a team of just 26 people on a$20 million budget. They're claiming it's the most capable open weight model ever released by a non Chinese company. The company's explicit goal is to give Western companies an alternative to Chinese models, something they can download, customize, and run on their own infrastructure. It's already become one of the top models used with OpenClaw. That's it for today. See you tomorrow.