AI models have surpassed human security researchers at finding vulnerabilities. Claude Mythos Preview found thousands of high-severity zero-day vulnerabilities across every major operating system and web browser, exploits that survived decades of human review. Anthropic is giving major technology companies a window to push patches before Mythos releases publicly, an effort dubbed Project Glasswing. For enterprises, this means faster-moving exploits at every layer of the stack: operating systems, browsers, server runtimes, npm dependencies. Organizations need to roll out patches faster than ever before.
What is Project Glasswing?
Project Glasswing is a consortium of major technology organizations, including AWS, Apple, Google, Microsoft, Nvidia, Cisco, CrowdStrike, JPMorganChase, Palo Alto Networks, and the Linux Foundation. Together, they are patching vulnerabilities before the model reaches the general public. Anthropic is committing up to $100M in usage credits for partners, with ongoing access afterward at $25 per million input tokens and $125 per million output tokens. It's also donating $4M directly to open-source security foundations (Alpha-Omega, OpenSSF, and the Apache Software Foundation) to finance their own efforts.
Claude Mythos Preview found thousands of zero-day vulnerabilities in major operating systems and browsers, with some bugs having survived decades of human review and millions of automated security tests.
What do Glasswing and Mythos mean for enterprises?
The volume of AI-discovered CVEs is about to increase sharply. Mythos is the first model operating at this level, but it won't be the last, and organizations without an up-to-date SBOM will be slowest to respond. Other frontier labs will follow.
The window between vulnerability discovery and exploitation is now hours, not weeks. Once a CVE drops, attackers can use coding agents to engineer an exploit within hours. Previously, enterprises shipped patches quarterly. Now they must ship daily. That means investing in vulnerability detection and prioritization tooling (Snyk, Wiz) that surfaces new CVEs as they drop. It also means rethinking package management (e.g. Flox) so that dependencies can be pinned and altered deterministically without lengthy scanning, which is a core discipline within software supply chain security.
MCP governance matters more than ever
A compromised MCP server exfiltrates sensitive data through tool calls. You need to trust certain MCP servers to get work done, but you also need guardrails around agents so any attack vector is stopped before it spreads.
AI agents often have broad access to internal systems (e.g. databases, file storage, APIs), and a single compromised tool can expose the entire system. To protect against this, limit agent access to only what is necessary, following the least-privilege principle. Audit every agent action to scope damage when a breach occurs. That means observability into the tools that an agent is calling, including all of the inputs and outputs.
You need a control plane that enforces permissions, logs every tool invocation, and gives security teams visibility into shadow MCP connections that developers may have spun up without review. Runlayer is built for this.
What is next for Claude Mythos?
Mythos will reach the general public soon. That means preparing for escalating CVEs, next-day patch urgency, and attack vectors across every layer of the stack. At the same time, Claude Mythos demonstrates what agentic AI can do at scale, and the opportunity for
