Anthropic unveiled Project Glasswing in early April, wrapping its new Claude Mythos Preview model behind a closed‑door program. Rather than a public blog or open waitlist, the company invited roughly 40 partners—primarily cybersecurity firms, cloud providers and critical‑infrastructure operators—to test the system. Anthropic told the broader market the model’s ability to autonomously discover zero‑day vulnerabilities made unrestricted release too risky.

Anthropic’s own Glasswing brief claims Mythos crossed a capability threshold unseen in commercial AI: real‑time, automated exploitation of previously unknown software flaws. The company priced access at $25 per million input tokens and $125 per million output tokens, pledging up to $100 million in credits for participants. Partners such as AWS and Microsoft echoed the security‑first narrative, framing early access as a defensive advantage.

Critics argue the restrictive rollout may mask more mundane constraints—compute costs, infrastructure limits and the economics of running a frontier model at scale. Whether safety concerns or budget pressures drive the gatekeeping, the episode forces the industry to confront how powerful AI tools are released. Anthropic’s approach shows that controlling access can become a strategic lever as much as a safety measure.