Good morning, I'm your AI Brief anchor. Here's what's happening in AI today, Wednesday, April 8, 2026.
Anthropic's Very Bad, Very Good Day
It’s a big day for news out of AI safety leader Anthropic, and it’s a mixed bag. The company is facing a wave of criticism from developers over its flagship model, Claude, while simultaneously rolling out major new safety and transparency initiatives.
The top story today is the growing frustration in the developer community. Reports are flooding platforms like GitHub and Hacker News from coders who claim recent updates have made Claude "unusable" for complex programming tasks. Developers say the model, once praised for its coding prowess, is now struggling with logic, generating buggy code, and failing on tasks it previously handled with ease.
The complaints point to a significant decline in performance, with some longtime users speculating that Anthropic may have implemented overly aggressive safety filters or altered the model in a way that inadvertently hobbled its coding capabilities. This backlash is a major headache for Anthropic, as developer trust is critical in the highly competitive world of foundational models. The company has yet to issue a formal response to the widespread complaints.
Project Glasswing Aims to Secure AI's Foundations
Moving on, while facing heat on its current product, Anthropic is looking far into the future of AI safety with a brand new initiative. The company has unveiled Project Glasswing, an ambitious effort to secure the very software that underpins powerful AI models.
Instead of just focusing on the behavior of the AI model itself, Glasswing targets the foundational libraries and code that AI systems are built on. The project’s goal is to use a technique called "formal verification." Think of it as mathematically proving that a piece of software works exactly as intended, with no hidden bugs or security vulnerabilities.
Anthropic argues that as AI models become more powerful and autonomous, we can't afford to have them running on a foundation that might crack under pressure. By securing these critical components, Project Glasswing aims to build a safer, more reliable ecosystem for all future AI development, preventing potential catastrophic failures before they can happen.
A Look Inside 'Claude Mythos'
And finally, in another move toward transparency, Anthropic has given the public a rare peek behind its safety curtain. The company released a "system card" for an unreleased internal test model named 'Claude Mythos.'
System cards are detailed documents that outline a model's capabilities, limitations, and the results of rigorous safety tests. The 'Claude Mythos' document reveals how Anthropic stress-tests its models for risks like self-replication, persuasion, and vulnerability to manipulation. For example, it details "red-teaming" exercises where testers tried to trick the model into revealing sensitive information or performing dangerous tasks.