Following Anthropic's announcement of Mythos finding zero-days, independent testing revealed a surprising finding: much smaller, cheaper open-weight models replicated much of the same security analysis. Researchers found that many smaller models successfully detected the core logic of sophisticated exploits, including the flagship FreeBSD vulnerability showcased by the frontier model.

This capability test suggests that AI cybersecurity prowess scales unevenly, contradicting the notion that only massive models can perform deep security reasoning. Eight different models, one costing just $0.11 per million tokens, successfully identified the flagship exploit chain. This uneven scaling means the capability frontier for security tasks is surprisingly jagged, not smooth.

The practical implication is that the defensible advantage—the "moat"—in AI security work resides in the orchestration system, not raw model intelligence. Since smaller models suffice for core detection, organizations can prioritize broad coverage and cost efficiency over deploying the most expensive frontier models for every step of the pipeline.

This research directly challenges the monolithic view of AI security development. The value proposition shifts toward embedding deep security expertise within the scaffolding that guides cheaper models through discovery, validation, and patch generation, rather than relying solely on a single, costly intelligence source.