Anthropic: Claude Mythos & Project Glasswing
Anthropic announces Project Glasswing, a program to restrict access to their new Claude Mythos model to vetted security researchers, aiming to use AI to find and fix critical software vulnerabilities.
Anthropic published the system card for Claude Mythos Preview, detailing the model's capabilities, safety evaluations, and risk mitigations for its cybersecurity-focused AI.
Anthropic's red team shares their assessment of Claude Mythos Preview's cybersecurity abilities, evaluating how the model performs on vulnerability detection and exploit analysis tasks.
Financial Times · www.ft.com
Anthropic made Claude Mythos available to Amazon, Microsoft, and Apple for detecting hidden software vulnerabilities, rolling out the model days after a source code leak.
Simon Willison · simonwillison.net
Simon Willison discusses Anthropic's decision not to broadly release Claude Mythos, instead restricting it to security researchers under Project Glasswing, and argues this cautious approach is warranted.
AI Models, Agents & Tools
Chinese AI lab Z.ai released GLM-5.1, a massive 754B-parameter MIT-licensed model focused on long-horizon tasks, available on Hugging Face at 1.51TB.
Google open-sourced Scion, an experimental testbed for orchestrating AI agents, providing researchers and developers with tooling to build and evaluate multi-agent systems.
The operator of acme.com reports that LLM scraper bots are overwhelming their HTTPS server, illustrating the growing infrastructure burden that AI training crawlers impose on smaller websites.