Too dangerous or just too expensive? The real reason Anthropic is hiding Mythos
3 hours ago
- #Compute Constraints
- #Claude Mythos
- #AI Safety
- Anthropic launched Claude Mythos Preview through Project Glasswing in April 2026, restricting access to about 40 vetted organizations, citing security risks due to its autonomous zero-day vulnerability exploitation capability.
- Anthropic's security rationale is supported by internal documents, partner endorsements, and the structured Glasswing consortium, positioning restricted release as a safety measure to give defenders a head start.
- Parallel evidence suggests compute constraints also influenced the restricted release, including Anthropic's compute expansion deals, capacity admissions, hiring signals, and a leaked draft mentioning high operational costs.
- The model's technical specifications (1 million token context window) imply high infrastructure strain, making a gated rollout sensible for both security and economic reasons.
- Anthropic likely emphasizes the security narrative over compute constraints for strategic advantage, as it frames the company as responsible and counters competitive attacks.
- The most plausible conclusion is that both security and compute factors are real and interacting, leading to a deliberate, restricted access model for Mythos Preview.
- The situation previews future AI deployment challenges where safety, economics, and technical constraints intersect, requiring balanced access models like Glasswing.