How AgentGateway Uses RFC 8693 Token Exchange to Secure AI Agent Identity

·4 views

In agentic AI systems, passing a user's broad-scoped identity token across multiple downstream services creates serious security risks, including over-privileged access and potential impersonation. RFC 8693, the OAuth 2.0 Token Exchange standard, addresses this by allowing a gateway to swap the original user token for a fresh, scoped, short-lived credential tailored to each downstream service. AgentGateway, a Rust-based AI proxy from the AI Agent Infrastructure Foundation, implements this pattern via its built-in backendAuth.oauthTokenExchange policy. The setup involves authenticating users through Keycloak, after which AgentGateway handles all token exchanges so upstream services never receive the original credential. The tutorial also covers token caching for performance, delegation with actor claims, and support for additional grant types such as JWT Bearer and Microsoft Entra On-Behalf-Of flows.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer Compares Spring Boot and Quarkus After Building a Microservices System

A seasoned Java developer with extensive Spring Boot experience built a small microservices system using Quarkus to evaluate the framework firsthand. The project comprised an identity service, catalog service, order service, and API gateway, all backed by a PostgreSQL instance, containerized via Docker Compose, and secured with JWT. A key difference observed was Quarkus's build-time annotation processing, which minimizes runtime reflection and delivers significantly faster startup and hot-reload speeds compared to Spring Boot's runtime evaluation model. For data access, Quarkus offers Panache with both active-record and repository patterns, whereas Spring Data JPA relies solely on the repository pattern with generated implementations. The developer concluded that while Quarkus shares surface-level similarities with Spring Boot, it represents a genuinely different philosophy toward Java framework design rather than simply being a lighter alternative.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Fusuma turns plain Markdown files into slides, PDFs, and OG images via CLI

Fusuma is an open-source, Node.js-based command-line tool that converts plain Markdown files into browser-ready slide presentations. Developers author content in standard Markdown, using triple-dash separators for slide breaks and HTML comments as layout directives, keeping files fully readable on platforms like GitHub. A built-in dev server with hot reloading, along with commands for PDF export and GitHub Pages deployment, all draw from the same single source file. The tool uses webpack 5 under the hood, supports Prism-based syntax highlighting, and can auto-generate Open Graph social card images when a git remote is configured. Its minimal configuration model and familiar Markdown authoring make it a lightweight alternative to proprietary presentation software.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Logically Inconsistent Prompts Can Force AI Reasoning Models Into DoS-Like Loops

Researchers from Zhejiang University and Alibaba, presenting at ICML 2026, found that deliberately corrupted prompts can trap AI reasoning models in long, unproductive internal loops. The attack works by feeding models logically inconsistent problems whose premises cannot be reconciled, causing them to generate far more tokens than usual. Using an evolutionary algorithm, the team refined prompts across five generations to maximise output length, testing successfully against DeepSeek-R1, Qwen3-Thinking, GPT-o3, and Gemini 2.5 Flash. In the most extreme case, response length on DeepSeek-R1 ballooned to 26 times that of normal outputs. Because the technique requires only external API access, it poses a realistic denial-of-service threat to commercial AI services, driving up compute costs and degrading performance for legitimate users.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MCP Server 'Nervous System' Adds Safety Rules and Audit Trails to Autonomous AI Agents

A new open-source tool called the Nervous System is an MCP server designed to impose enforceable behavioral constraints on autonomous LLM agents. It addresses common failure modes such as lost session context, unauthorized file edits, and irreversible actions by providing mechanically enforced rules rather than mere prompt-based suggestions. Key features include preflight checks before file edits, session handoff logging, a hash-chained audit trail, and a kill switch for destructive actions. A read-only public package is available via npm and can be added to any MCP client, while a hosted version with a fuller toolset has reportedly been running in production for several months. Individual audit functions are also accessible on a pay-per-call basis using USDC on Base for CI and agent-to-agent pipeline use cases.

0 comments Read more at DEV Community