Anthropic's Glasswing Update Pushes Secure Agents

Anthropic's May 22 Project Glasswing update is the most developer-relevant item in its recent feed. The company says Mythos Preview has found more than ten thousand high- and critical-severity vulnerabilities across partner software, and it is now sharing a toolkit for security teams that want to run similar agentic workflows. This is not just a research milestone. It signals that Anthropic is turning model capability into operational security infrastructure, with scanning, triage, reporting, and disclosure steps wrapped into a repeatable workflow.

Key Features or Updates

Glasswing's initial update centers on Mythos Preview, Anthropic's security-focused model snapshot. Anthropic says partners have used it to find large numbers of serious vulnerabilities, and the company is exposing a release that includes skills, a harness for mapping codebases and spinning up scanning subagents, and a threat model builder. It also expands the coordinated vulnerability disclosure program and shows the operational side of agentic security work.

Impact on Developers

For developers and security teams, the important shift is from one-off AI findings to a structured workflow that can scale across many projects. The update points to a future where AI helps find issues faster, but humans still control verification, disclosure, and patching. That makes the tooling more relevant for teams that need reproducible, auditable security automation instead of ad hoc prompt sessions.

How to use it

If your team does defensive security work, treat this as a preview of the workflow shape to watch for in Claude tooling. The practical next step is to evaluate how a harness, scanning subagents, and structured triage could slot into your existing vulnerability disclosure or code review pipeline. If Anthropic opens access for qualified customers, security teams should test it against their own internal repos and patch workflow before depending on it.