AWS AgentCore adds secure OAuth MCP integration

TLDR: AWS is quietly tightening the agent infra story while security folks ship both new attack maps and a local-only AI pentest assistant.

AWS connects AgentCore Gateway to OAuth-protected MCP servers

Amazon Bedrock AgentCore Gateway now supports connecting to OAuth-protected Model Context Protocol (MCP) servers using the Authorization Code flow, with AWS walking through a full configuration example as of 2026-04-07. The guide covers the AgentCore Gateway setup, the OAuth client registration, and how to route agent tool calls through a centralized gateway into protected internal services.

If you are trying to standardize agent access to internal tools across teams, this matters. You get a single control plane for how agents authenticate to MCP servers instead of sprinkling OAuth logic across every agent runtime. It also nudges you toward treating tools like first-class services with proper identity, logging, and policy instead of ad hoc shell access.

The practical next step is to align this with your existing IdP and secrets management so you can rotate credentials without redeploying agents.

Read more →

---

AWS shows RL-based tool calling with serverless SageMaker

Amazon SageMaker AI published a recipe for fine tuning Qwen 2.5 7B Instruct for tool calling using RLVR on serverless infrastructure as of 2026-04-07. The post walks through building datasets for three distinct agent behaviors, defining a tiered reward function, training configuration, evaluation on unseen tools, and serverless deployment.

If you are unhappy with vanilla tool-use behavior from off-the-shelf models, this is a concrete blueprint. The detail on reward shaping and evaluation is useful for any team trying to get agents to call tools more reliably without hand curating prompts. Because it runs on serverless SageMaker, you can experiment without long-lived GPU clusters, although cost and latency still need to be measured in your environment.

Expect more patterns like this: narrow RL recipes targeting specific agent skills instead of full-stack custom models.

Read more →

---

METATRON ships a local-only AI penetration testing assistant

METATRON is a new command line penetration testing assistant for Parrot OS and other Debian-based Linux distributions that combines automated recon tooling with a locally hosted large language model, as of 2026-04-07. You feed it a target IP address or domain and it orchestrates familiar tools like nmap for port scans and nikto for web server checks, with analysis done entirely on your machine.

For anyone doing security work on sensitive networks, the local-only design and lack of required API keys reduce the data leakage risk that comes with cloud LLMs. It is written in Python 3 and focused on automation of existing toolchains rather than flashy UI, so it should slot into current workflows. The tradeoff is that capability is bounded by your local hardware and the chosen model.

This is also a solid reference pattern for other privacy sensitive agentic workflows where cloud LLMs are a non starter.

Read more →

---

Quick Hits

• Build AI-powered employee onboarding agents with Amazon Quick Amazon shows how to wire Quick into HR systems so an onboarding agent can answer policy questions and track paperwork. Useful if you are piloting narrow internal agents before tackling broader workflows.

• Microsoft's Agent Stack Confuses Developers While Rivals Simplify Forbes argues Microsoft has an overloaded agent story while Google Cloud offers a cleaner stack through the Agent Development Kit, Agent Engine, and Vertex AI Agent Builder. Worth a skim if you are picking a managed agent platform.

• Google DeepMind Researchers Map Web Attacks Against AI Agents DeepMind categorizes behavioral control traps and systemic traps that target web-connected agents, including jailbreaks in external content and inter agent dynamics. Good input for your threat models before you let agents browse or call each other.

• Lula: multi-agent coding assistant with sandboxed Rust engine Show HN multi agent coding assistant that executes code in a sandboxed Rust environment. Interesting if you want inspiration for safe execution layers for autonomous coding agents.

• jmux: tmux-based environment for humans and coding agents A tmux workflow aimed at juggling 5 to 10 Claude Code sessions plus servers and logs without moving to a GUI orchestrator. If you live in terminals and run many agents side by side, this is worth testing.

• Freestyle: sandboxes for coding agents Launch HN startup building a cloud specifically for coding agents, starting from SQL helpers and moving to full app builds and deploys. The high Hacker News engagement suggests real demand for hosted sandboxes with opinionated agent tooling.

• [AINews] Gemma 4 crosses 2 million downloads Latent Space notes Gemma 4's rapid adoption and how it is becoming a default open model choice. If you want a widely supported base for local agents, this signals ecosystem maturity.

• Google AI Edge Gallery Simon Willison reviews Google's iOS app for running Gemma 3 and Gemma 4 variants locally on iPhones, calling it fast and useful. Mobile on device agents are becoming much more practical.

• OpenAI Safety Fellowship OpenAI launches a pilot fellowship to fund independent safety and alignment work and build talent. Relevant if you want to do safety research without joining a big lab.

• scan-for-secrets 0.3 Small Python tool to scan files for secrets, now with a redact option and a redact_file helper. Handy pre commit step before you feed repos to agents or share logs.