Anthropic Integrates Project Glasswing Into Tiered Access for Advanced Cyber LLMs

Anthropic disclosed that its Project Glasswing initiative has been folded into a new tiered-access framework for the company's suite of cyber‑focused large language models, including the Opus, Sonnet and Mythos systems. The move consolidates the program that previously offered a separate pathway for security researchers, aligning it with the broader access structure used for Anthropic's most capable AI models.

Project Glasswing was originally launched as a sandbox for vetted defenders to experiment with AI‑driven threat detection and response tools. By merging the effort into a tiered system, Anthropic aims to streamline onboarding while still preserving a controlled environment for those with proven expertise in cybersecurity.

The tiered model categorizes users based on their verification status and intended use cases, granting progressively higher levels of model capability. At the top tier, users can interact with Opus, Sonnet and Mythos—models that incorporate the latest advances in language understanding, reasoning and code analysis, and that have been fine‑tuned for security‑related tasks.

One notable change for the vetted defender cohort is the reduction of Claude guardrails within the models they receive. Guardrails are built‑in safety constraints designed to limit risky or undesirable outputs. By offering versions with fewer of these restrictions, Anthropic hopes to give experienced security professionals greater latitude to probe edge cases, simulate adversarial behavior and uncover hidden vulnerabilities.

The adjustment has sparked discussion in the cybersecurity community about balancing openness with safety. While greater flexibility can accelerate research and improve defensive tooling, it also raises questions about potential misuse. Anthropic has indicated that monitoring and revocation mechanisms remain in place, and that future iterations of the tiered program may incorporate additional safeguards based on feedback from early participants.