Anthropic has released an initial update on Project Glasswing, its joint initiative with around 50 partners to secure critical global software systems before advanced AI models can be misused against them. Using the Claude Mythos Preview model, participants have identified more than 10,000 high- or critical-severity vulnerabilities across essential software platforms. Partners such as Cloudflare, Mozilla, and Oracle reported major increases in bug detection and patching speed, while external testers confirmed Mythos Preview’s superior performance on multiple cybersecurity benchmarks.

Anthropic also used Mythos Preview to scan over 1,000 open-source projects, uncovering more than 6,000 high- or critical-severity vulnerabilities, with independent verification confirming over 90% as valid. The company noted that the main challenge now lies in verifying and patching the large number of vulnerabilities, as human triage and disclosure processes lag behind AI discovery rates. Some maintainers have requested slower disclosure to manage patch workloads.

To support defenders, Anthropic has launched tools such as Claude Security and the Cyber Verification Program, and partnered with the Open Source Security Foundation to help maintainers process reports. The company emphasized that Mythos-class models will not be publicly released until stronger safeguards are developed.