Claude Mythos is live on Amazon Bedrock. Sort of.
It's a gated research preview — meaning you can't just sign up and start using it. Access is limited to what AWS calls "allowlisted organizations," with Anthropic and AWS prioritizing internet-critical companies and open source maintainers. The program is called Project Glasswing, and it's not for general use yet.
What makes Mythos different from Anthropic's other models is the focus. This one is built specifically for cybersecurity work: identifying vulnerabilities in software, analyzing large codebases, and complex security reasoning. Anthropic is pitching it as a tool for security teams to find and fix issues before they become incidents. Whether it actually delivers on that is hard to evaluate when almost nobody can access it, but the direction is interesting — a model class purpose-built for a specific high-stakes domain rather than a general assistant with security added on top.
AWS Agent Registry: MCP server included
The other headline this week is AWS Agent Registry, which launched in preview through Amazon Bedrock AgentCore.
The idea is straightforward: as organizations build more AI agents, they end up with a sprawl problem. Teams duplicate tools. Nobody knows what already exists. Agent Registry is meant to be the internal catalog that fixes that — a searchable directory of agents, tools, skills, and MCP servers that teams can discover and reuse.
What caught my attention is that the registry itself is accessible as an MCP server. You can query it directly from your IDE, which makes the discovery workflow a lot more practical than navigating another console. It also ships with approval workflows and CloudTrail audit trails, so there's governance built in from the start.
For anyone working in the MCP space, this is worth watching. AWS is essentially treating MCP servers as first-class citizens in their agent infrastructure catalog.
Other stuff from this week
S3 Files — Amazon S3 now supports mounting buckets as file systems, built on EFS technology. The pitch is that your applications can access the same S3 data through both file system APIs and the S3 API without changing code. Multi-terabyte per second read throughput, actively used data cached. If you've ever had to choose between EFS and S3 for a workload, this is interesting.
OpenSearch + Managed Prometheus — OpenSearch Service added native Prometheus integration with direct PromQL support, plus OpenTelemetry GenAI semantic convention support for tracing LLM execution. The agent tracing part is the relevant bit if you're running AI infrastructure — you can now correlate slow traces back to logs and overlay Prometheus metrics in one place instead of jumping between tools.
Bedrock IAM cost allocation — You can now tag IAM principals with attributes like team or cost center, and that data flows into Cost Explorer and the detailed Cost and Usage Report. Useful if you're trying to track which teams are actually spending what on model inference, especially as agent workloads scale up.
Rigetti Cepheus on Braket — Amazon Braket added Rigetti's 108-qubit Cepheus QPU, which is the first 100+ qubit superconducting processor on the platform. Supports Braket SDK, Qiskit, CUDA-Q, and Pennylane. Niche, but notable if you're in the quantum computing space.
Quick take
The Mythos announcement is the one I'm most curious about. Cybersecurity is an interesting choice for a purpose-built model — it's a domain where hallucinations have real consequences, so the bar for reliability is higher than most use cases. The gating makes sense given that. What Anthropic decides to do with access over the next few months will probably tell us more about where this category is going than the launch announcement does.
Agent Registry is the more immediately practical release. If your team is building with agents, a centralized catalog with MCP server access and audit trails is the kind of boring infrastructure that actually matters.
Top comments (0)