Anthropic won't release its newest model, Mythos Preview, to the public. The reason is simple: it's too good at finding security flaws like a Nation-State. The model can spot vulnerabilities and chain them into complex exploits at a level that alarmed even its own creators. Instead of a public launch, Anthropic is sharing Mythos with a small group of tech companies so they can patch their defenses first. What makes Mythos different from existing security scanners is how it understands code. Traditional tools check files one at a time for known patterns. Mythos reads entire codebases at once, tracking how data flows between modules. It can spot race conditions and logic bugs that span multiple files, then generate proof-of-concept exploits and test them in sandboxes to confirm they work. Logan Graham, who leads offensive cyber research at Anthropic, told NBC News the model doesn't just find individual vulnerabilities. It chains them into complicated attacks. The worry extends beyond Anthropic's walls. Treasury Secretary Scott Bessent convened a meeting with major financial institutions this week to discuss what AI means for economic stability. Katie Moussouris, CEO of Luta Security, predicted we'll see major outages with cascading effects, similar to when cloud providers go down. Casey Ellis, founder of Bugcrowd, put the problem bluntly: defenders need to be right every time, but attackers only need to succeed once. Graham expects competitors, including those in China, will release comparable models within six to twelve months. "That's a pretty crazy time frame," he told NBC News. "Usually preparations for things like this take many years." The models we have today are already finding real vulnerabilities. Claude, Codex, and Gemini spot subtle flaws in open-source code that experienced maintainers miss. Cynthia Kaiser, a former senior FBI cyber official now at Halcyon, worries most about "the wannabes," people who lacked the skill to launch serious attacks a year ago but now have powerful AI tools. Hospitals and manufacturing plants will remain top targets, she said, because attackers go where downtime hurts most.