In the first weeks of March, Omar Santos got early access to the model that became Anthropic’s Mythos. About an hour of testing was enough to convince him cybersecurity had changed for good.
Omar, a Distinguished Engineer in AI Security Engineering at Cisco and co-chair of the Coalition for Secure AI, says his bigger worry is how fast everyone else is catching up. Open weight models are closing in on the frontier labs, so attackers no longer need access to Mythos to find and exploit vulnerabilities.
KB and Omar get into why Cisco won’t send a finding to engineering without proof of an active exploit, how a public open source fix gives attackers a ready-made recipe, and why a library a student abandoned years ago could be running inside your patient records system. Omar wants defenders to go further than finding and fixing faster. That means refactoring code, adding compensating controls, and turning the risk register from an annual exercise into something that runs continuously.
About Omar:
Omar Santos is a Distinguished Engineer at Cisco, focused on AI security, cybersecurity research, incident response and vulnerability disclosure. He co-chairs the Coalition for Secure AI (CoSAI) and the FIRST PSIRT Special Interest Group, sits on the board of OASIS Open, and chairs the OpenEoX and CSAF technical committees, where his work led to CSAF becoming an ISO standard. He also led the DEF CON Red Team Village for several years.
Omar has written more than 25 books, 21 video courses and over 50 research papers, and holds multiple patents. Before Cisco, he served in the United States Marines.
Keywords: Mythos, Anthropic, Project Glasswing, frontier AI, open weight models, AI security, AI in cybersecurity, vulnerability management, zero days, exploit development, open source security, software supply chain, SBOM, OpenEoX, CSAF, CodeGuard, Coalition for Secure AI, CoSAI, Cisco, compensating controls, eBPF, secure by design, zero trust, incident response, threat hunting, continuous assessment, risk register, GRC, CISO, board, KBKast
