- Anthropic's Mythos AI model can generate working exploits for browser sandboxes 72.4% of the time, up from under 1% for previous models, threatening the foundational security of the internet.
- Sandboxing is a critical cybersecurity layer in browsers, apps, and cloud computing, and its compromise could allow malicious code to take control of devices or cloud infrastructure.
- Smaller models are rapidly gaining similar capabilities, making widespread exploitation likely even if Mythos is not broadly released.
- The current plan to share the model only with select cybersecurity professionals may be insufficient due to the vast number of open-source and commercial software at risk.
- This represents a shift where AI achieves superhuman capability in finding vulnerabilities, potentially leading to widespread disruption if not addressed.