Anthropic's AI security model Mythos was tested on the widely-used curl codebase. After scanning 176,000 lines of code, it initially flagged 5 potential vulnerabilities. Manual review by the curl security team determined most were false positives or regular bugs, leaving only one low-severity issue confirmed. Curl's founder noted the codebase had already been refined by many tools and audits, making new, significant finds difficult. While AI tools offer advantages in understanding protocols and code context, they currently excel at finding known error types, not novel logic flaws. The test highlights that rigorous security engineering remains foundational for core software safety.