The Claude Mythos Preview AI model is the result of a partnership between AWS and Anthropic as part of Project Glasswing. This special AI tool will be released in a limited number of copies to help make the world's most important software systems safer. AWS has already used machine learning a lot in its security measures for its infrastructure.
AWS was able to stop more than 300 million malicious S3 encryption attempts in 2025 alone by looking at more than 400 trillion network flows every day. However, because AI models can create working exploits, AWS and Anthropic are being careful about how they share their work. You can now get it through Amazon Bedrock in a gated research preview.
Only a small number of organizations can access it. These organizations are mostly internet-critical companies and major open-source maintainers, whose software affects millions of users around the world. Organizations that test the model get strict enterprise-grade security measures that keep their production assets safe from unnecessary risks.
Customer-managed data encryption lets teams safely look into AI's capabilities without giving up their privacy. Strict Virtual Private Cloud (VPC) isolation keeps workloads safe by keeping them separate from other resources and threats from outside. Automated Reasoning safeguards make sure that there are no factual hallucinations 99% of the time. The new AWS Security Agent has officially made its ability to do enterprise penetration testing available to everyone.
This tool changes the process from a manual, periodic bottleneck to a constant, on-demand capability that works all the time.
It finds and checks for weaknesses on its own in different cloud environments, such as AWS, Azure, GCP, and on-premises systems. The Amazon Security Agent is different from regular security scanners in that it tries to take advantage of weaknesses by using targeted attack chains.
