Anthropic limits Mythos model
anthropic restricted access to its claude mythos model after internal tests found it could autonomously exploit thousands of zero-day vulnerabilities.
the article examines the geopolitical security implications of anthropic's claude mythos, a model with advanced cyber-offensive capabilities that anthropic deemed too dangerous for public release. testing revealed mythos can autonomously execute multi-stage cyberattacks, uncovering thousands of vulnerabilities in major operating systems and crashing essential infrastructure. anthropic mitigated immediate risks by restricting access to staff and project glasswing partners, a defensive initiative designed to patch vulnerabilities before deployment. however, the author warns that mythos lowers the barrier for ai-enabled coups by reducing the coalition size needed to hack autonomous military systems and disrupt governance.