Anthropic's Calculated Sacrifice: Why It Capped Its Most Powerful Security AI

Anthropic intentionally limited Opus 4.7's public capabilities because Claude Mythos had become too effective at discovering zero-day vulnerabilities—finding exploits that evaded security researchers for decades in hardened systems like OpenBSD. This decision marks a critical inflection point in AI development: when systems become powerful enough to discover real-world exploits and chain vulnerabilities into working attacks, responsible deployment may require restricting access rather than maximizing performance. The move signals that the industry must rethink how it manages AI systems dangerous enough to break what humanity depends on.