Details
- Anthropic is expanding Project Glasswing from roughly 50 initial partners to about 200 total, adding approximately 150 new organizations that must meet security requirements before accessing its Claude Mythos Preview model.
- The expanded cohort spans more than 15 countries and now more fully covers critical infrastructure sectors including power, water, healthcare, communications, and hardware, as well as major open-source vendors whose code underpins government and commercial systems.
- Partners are using Claude Mythos Preview to scan large codebases, having already surfaced over 10,000 high- or critical-severity vulnerabilities, and increasingly rely on the model not just to find flaws but also to generate patches and run pre-release security checks.
- Anthropic is complementing Mythos Preview with broader-access tools like Claude Security, built on public models such as Claude Opus 4.8, and is preparing to share internal vulnerability-scanning tooling with trusted security teams to help standardize AI-assisted defensive practices.
- Framing Mythos-class systems as a near-term industry norm, Anthropic argues that many AI developers may release similarly capable models without strong safeguards within 6–12 months, and is working with governments and the security community on scalable norms, disclosure practices, and a Cyber Verification Program to safely extend advanced cyber capabilities.
Impact
The expansion signals that AI-first vulnerability discovery at internet scale is moving from pilot to early deployment, especially across critical infrastructure and open-source software. Over the next 12–24 months, security teams may be pushed to re-architect workflows around rapid AI-driven finding, triage, and patching, while regulators and standards bodies confront the dual-use reality of frontier cyber models and the need for enforceable safeguard norms.
