Presented by
Good morning, AI enthusiasts. There's a scene in every heist film where the safecracker opens a vault that's beaten everyone else for decades. Claude Mythos just did the equivalent in software, finding a 27-year-old bug in one of computing's most hardened operating systems and a 16-year-old flaw in code that automated tools had scanned five million times without catching.
The model launched today alongside Project Glasswing, a defensive cybersecurity effort backed by 11 major tech partners.
In today's recap:
Anthropic's Claude Mythos hunts zero-days in production code
GLM-5.1 matches Claude Opus 4.6 at a fraction of the cost
Audit your codebase for zero-day vulnerabilities with Claude Code
Meta open-sources versions of its next AI model family
4 new AI tools, prompts, and more
ANTHROPIC
Anthropic unveils AI that hunts real-world zero-days
Recaply: Anthropic just launched Project Glasswing, a cybersecurity initiative with 11 tech partners including AWS, Microsoft, and Google. The goal: use Claude Mythos Preview to find and fix zero-day vulnerabilities at scale before attackers do.
Key details:
Mythos Preview scans code on its own, finding and chaining flaws without human direction. It has already flagged vulnerabilities in every major operating system and browser.
Mythos Preview scores 83.1% on CyberGym vs. Opus 4.6's 66.6%, and 77.8% on SWE-bench Pro vs. 53.4%. Anthropic is committing $100M in usage credits for defensive work.
The model found a 27-year-old flaw in OpenBSD that let any attacker remotely crash a machine just by connecting to it, and a 16-year-old FFmpeg bug that automated tools had scanned five million times without catching.
Over 40 organizations building critical software already have access. Anthropic is also donating $4M to open-source security groups.
Why it matters: Cybersecurity has always been uneven. Attackers need to find one flaw. Defenders must catch them all. Mythos Preview found bugs that survived decades of expert review and millions of automated tests. With 11 major partners using it on day one, Anthropic is betting the only path forward is putting this in defenders' hands first, before it spreads.
PRESENTED BY HUBSPOT
How can AI power your income?
Ready to transform artificial intelligence from a buzzword into your personal revenue generator
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
ZAI
New open-source model matches Opus 4.6 performance
Recaply: Zai just released GLM-5.1, an open-source model that matches Claude Opus 4.6 on agentic coding benchmarks and goes further, sustaining optimization over 8 hours and 6,000 tool calls when given enough time.
Key details:
GLM-5.1 breaks complex tasks down, runs experiments, and revises strategies through repeated iteration, staying productive over longer sessions where previous models plateau after 50 tool calls.
GLM-5.1 scores 58.4% on SWE-bench Pro vs. Opus 4.6 at 57.3%, though Opus 4.6 still leads on KernelBench Level 3 at 4.2x speedup vs. GLM-5.1's 3.6x.
Released under an MIT License and compatible with Claude Code and OpenClaw, GLM-5.1 lets builders swap in the model without changing their existing agentic workflows.
Why it matters: Open-source models have been closing the gap on frontier performance for years, but GLM-5.1 is the first to do it credibly on long-horizon agentic tasks, which is where production engineering work actually happens. It's MIT-licensed and API-compatible with Claude Code, so teams can test cost reduction without infrastructure changes. At roughly 1/3 the cost of Opus 4.6, the savings compound fast for any agent running thousands of tool calls per session.
TUTORIAL
Audit your codebase for zero-day vulnerabilities with Claude Code
Recaply: In this tutorial, you will learn how to scan your codebase for security flaws using Claude Code. You'll find common issues like SQL injection and XSS, plus complex logic errors that rule-based tools miss.
Step-by-step:
Open your terminal, go to your project directory, and run
claudeto start a Claude Code session in your repo's root.Type
/security-reviewto start a full scan. Claude reads your files, traces data flows, and checks how components interact. It doesn't just match known patterns.Review the results sorted by severity. Focus on High and Critical items first. Each finding shows the file, a code snippet, and explains the risk, for example SQL injection via unsanitized user input.
Ask Claude to fix a specific issue. For example: "Fix the SQL injection in
api/users.js." Claude proposes a patch, explains the change, and waits for your approval. Nothing is applied without it.Set up the GitHub Action at github.com/anthropics/claude-code-security-review to run scans on every pull request. It posts inline comments with flagged issues before any code gets merged.
Pro tip: Claude Code catches logic flaws and broken access control that tools like ESLint or Semgrep will never flag. Run it on older code first. The code that hasn't been touched in years is where the 27-year-old bugs live.
META
Meta plans to open source next AI model family
Recaply: Meta just announced it will release open-source versions of its next AI model family. It's the first family built under CEO of AI Alexandr Wang. But Meta's biggest models will stay closed. The company is moving to a hybrid approach.
Key details:
Meta will check for safety risks before releasing anything as open source. The models are built to reach people through WhatsApp, Facebook, and Instagram, not through enterprise contracts.
Wang joined Meta via a $15B deal with Scale AI last year. The new family is designed to close the gap after Llama 4 fell behind rivals like GPT-5 and Claude Opus.
According to Axios, Wang sees Anthropic and OpenAI as focused on governments and enterprise. Meta's open models are meant to be the U.S.-made choice for consumers and developers worldwide.
Meta expects its new models may not top every benchmark. Both OpenAI and Anthropic have hinted they have major releases coming soon.
Why it matters: Meta's open-source moves have always been more strategy than principle, and Wang makes that plain. The real signal is what stays closed: its biggest models. Alibaba just did the same thing with Qwen. A pattern is forming across the industry. Open models are a floor now, not a ceiling. How the new family stacks up against GPT-5 and Mythos is a different question, and Meta already knows it won't lead across the board.
TOOLS
Trending AI Tools
🛡️ Composio Protection - Security layer that checks every AI agent tool call against OAuth scopes, permission rules, and allowlists before it reaches your apps, with full audit logging
📦 WildDet3D - Allen AI's open-source model for 3D object detection
🚀 Rocket 1.0 - Business platform combining competitive intelligence, strategy, and production-grade building in one shared workspace and memory
⚙️ Superset - IDE for the AI agent era
NEWS
What Matters in AI Right Now?
OpenAI CEO Sam Altman revealed that Codex has hit 3 million weekly active users, resetting usage limits to celebrate and committing to do the same for every additional million users up to 10 million.
Jeff Bezos's secretive start-up Project Prometheus hired Kyle Kosic, an xAI co-founder who led infrastructure behind its Colossus supercomputer, as it pushes to build AI systems that operate in the physical world beyond chatbots and coding tools.
The AI Alliance launched Project Tapestry, an open-source platform for globally federated training of frontier AI models, with Turing Award winner Yann LeCun joining as Chief Science Advisor.
WorldLabs rolled out two model updates: Marble 1.1 with improved lighting, contrast, and fewer visual artifacts, and Marble 1.1-Plus, a new model built for larger and more complex 3D environments at scale.
DeepSeek rolled out a limited V4 gray release with a new mode switcher in its chat UI, adding Expert Mode and Vision Mode alongside the default Fast Mode.
X launched a new photo editor in its post composer with drawing, text, blur, and a Grok-powered "edit with words" feature for AI-driven edits, starting on iOS with Android coming soon.
Intel announced it is joining Terafab alongside SpaceX, xAI, and Tesla to refactor silicon fabrication technology, with the initiative targeting 1 TW per year of compute to power advances in AI and robotics.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
