Claude Mythos Preview Faces Open-Model Pushback

Anthropic’s demo still shows real cyber strength, but the report weakens the idea that Mythos alone created a new, unassailable frontier. The bigger lesson is that cybersecurity performance is jagged: model size, price, and vendor prestige do not translate smoothly across tasks.

• –Eight of eight tested models found the FreeBSD issue once the relevant snippet was isolated, including a 3.6B-active open model priced at $0.11 per million tokens.
• –The harder OpenBSD chain still separated models, but a 5.1B-active open model recovered the core reasoning, which undercuts any simple “only frontier models can do this” story.
• –The same models reshuffled rankings across tasks, with small open models beating many frontier models on a basic security-reasoning probe.
• –For builders, the moat looks more like the surrounding security system, prompts, triage, and workflow than the raw model alone.