AI #163: Mythos Quest
Anthropic's new AI found security flaws that could 'break the internet' but is being used responsibly.
Anthropic's newly revealed Claude Mythos AI model has reportedly discovered critical security vulnerabilities in every major operating system and web browser, vulnerabilities so severe that their immediate release could 'break the internet' and cause widespread chaos. According to a detailed analysis by Zvi on LessWrong, Anthropic chose not to exploit these flaws but instead launched Project Glasswing, a controlled initiative to make the Mythos model available to cybersecurity companies. The goal is to enable rapid patching of the world's critical software infrastructure before the vulnerabilities can be weaponized by malicious actors.
In other significant AI news, Google released Gemma 4, a new open model that could be the best in its weight class for local deployment on phones and computers, potentially enabling cost-free, local AI agent setups. Suno also received a major upgrade for AI song generation. Meanwhile, Anthropic's business metrics are soaring, with the company reporting $30 billion in annual recurring revenue (ARR), a massive jump from $9 billion at the start of 2026 and $19 billion just a month prior in February.
- Claude Mythos discovered critical vulnerabilities in all major OSes and browsers that could cause internet-wide chaos.
- Anthropic launched Project Glasswing to responsibly share Mythos with cybersecurity firms for patching, not exploitation.
- Anthropic's ARR skyrocketed to $30B, while Google's Gemma 4 could enable powerful local AI agents.
Why It Matters
This demonstrates both the immense power and profound responsibility of frontier AI, setting a critical precedent for security disclosure.