After conducting internal tests, Cloudflare is warning that Anthropic’s new AI model Mythos may be capable of autonomously chaining together harmless software flaws into severe cyberattacks.