Has Mythos just broken the deal that kept the internet safe?
5 hours ago
- #Cybersecurity
- #Internet Safety
- #AI Sandbox Exploits
- Mythos, a large AI model by Anthropic, can generate working exploits for browser sandboxes 72.4% of the time, a drastic increase from under 1% with previous models.
- The internet's security relies on sandboxes (e.g., for browsers, cloud VMs) to contain untrusted code, but AI advances threaten these defenses, potentially enabling widespread device takeovers or cloud breaches.
- Despite Mythos's size and high cost, its capabilities may quickly transfer to smaller models or be enabled by new hardware, making widespread exploitation feasible.
- Anthropic is cautiously releasing Mythos to select cybersecurity professionals for preemptive fixes, but this approach risks missing vulnerabilities in countless critical open-source and commercial software projects.
- The rapid progress in AI poses an unprecedented cybersecurity risk, potentially destabilizing internet security and requiring new defensive strategies beyond current measures.