A new report on Anthropic’s Mythos preview describes how advanced agentic AI capabilities exposed long-standing software flaws and elevated corporate governance and security expectations. The piece says Anthropic identified issues while testing Mythos, which it frames as a shift toward systems that can execute more complex tasks—raising risks of autonomous exploitation and multi-step attacks. In response, Anthropic launched Project Glasswing, described as a coalition providing restricted access to the U.S. Cybersecurity and Infrastructure Security Agency (CISA) and a group of U.S. corporates, including Microsoft, Apple, and J.P. Morgan, to identify and fix critical vulnerabilities before broader release. The story’s governance argument is that agentic AI should be treated as an autonomous agent system—not just a chatbot—requiring strict oversight, central monitoring, and vendor and security controls to prevent unverified or hostile code execution. For universities and research partners deploying or evaluating cutting-edge AI systems, the reporting adds pressure to strengthen AI governance policies, including security review processes and accountability mechanisms for tool usage.