May 2025 Summaries

The Model Context Protocol (MCP) is an open standard that enables AI models to connect with data sources and services in a standardized way. An MCP server acts as a bridge between the AI model and various tools, resources, or capabilities, providing a uniform interface for developers to access these resources. The server's architecture consists of several key components, including the communication layer, request handlers, context stores, session orchestrators, and caching layers. These components work together to manage sessions, handle requests, store data in context stores, and provide fast access to frequently used or expensive-to-fetch data through caching mechanisms. MCP servers help maintain a clean separation between the AI model's logic and the server's action and data retrieval capabilities, making it easier for developers to build maintainable and composable applications.

Free trial abuse is becoming increasingly sophisticated, with attackers using automated scripts to create fake accounts en masse and farm backend resources without contributing to revenue. The problem has evolved from simple tactics to industrialized attacks that bypass traditional defenses such as CAPTCHA, email verification, and IP-based rate limiting. The economics of modern SaaS have changed, with high-value inference operations embedded in the core user flow, making it more damaging. Traditional defenses are no longer effective, and attackers can now use open-source "abuse kits" to simulate clean browser environments and evade detection. To combat this, WorkOS Radar provides real-time behavioral analysis, detecting automation and identifying shared connections across accounts, while allowing for customization of defenses through its Actions system.

OpenAI's Codex, a cloud-based agent, aims to become your AI coworker by handling entire development workflows, marking the first major shift from AI assistance to AI delegation in software engineering. The `codex-1` model powers this agent, operating as a full development agent that handles tasks such as debugging multi-file issues, writing comprehensive tests, and implementing features across complex codebases. Unlike earlier models, Codex is trained using reinforcement learning on actual coding tasks, achieving 75% accuracy on OpenAI's internal software engineering benchmarks and strong performance on SWE-Bench Verified. The system runs in isolated cloud containers with no internet access, reads AGENTS.md files for project-aware intelligence, and executes end-to-end workflows that include implementing features, running test suites, fixing errors, and committing changes. This represents a fundamental change in how AI assists knowledge work, shifting from making humans faster at manual tasks to taking over entire workflows while keeping humans in control of decisions and review.

The Device Authorization Grant (Device Code Flow) is a practical solution for secure user authentication on devices without keyboards, such as smart TVs and IoT devices. It offloads the authentication process to a secondary device by sending a verification code to the user's phone or laptop, which they then enter, allowing the primary device to act on behalf of the user. The flow includes several key security features, such as short-lived codes, rate-limited polling, explicit user consent, and phishing resistance. However, it also has vulnerabilities that can be exploited by malicious actors, particularly in cases where users are tricked into entering maliciously generated user codes. To defend against these attacks, developers must educate users, restrict device code usage to high-trust applications, use conditional access and anomaly detection, monitor token usage patterns, and never assume implicit safety just because users land on a trusted domain.

MCP (Model-Client-Protocol) sampling flips the traditional flow of language model requests by allowing servers to initiate requests to clients for completions, creating a more powerful and programmable pattern for using AI inside systems. This protocol enables structured, repeatable AI use with human oversight, bringing transparency, controllability, and audibility to workflows. Key benefits include server-initiated reasoning, human review, and structured, repeatable AI use. MCP sampling works by having the server send a request to the client, which acts as an intermediary between the server, user, and model, allowing humans to preview, edit, and approve requests before they are sent to the model. This process enables automated workflows with manual checkpoints, reproducible, auditable AI decisions, and human + AI collaboration on a per-task basis. MCP sampling is suitable for use cases such as content moderation, customer support, decision-making systems, structured data extraction, and form completion, offering a scalable way to combine AI speed with human judgment without bottlenecks or blind trust. By giving servers the power to think and ask, humans the ability to observe and approve, and teams the ability to build AI flows that are scalable, transparent, and accountable, MCP sampling transforms how AI interacts with systems.

A comprehensive guide to email deliverability troubleshooting and best practices is presented. The guide outlines the steps to diagnose and fix common issues that prevent emails from reaching inboxes, including checking DNS records (SPF, DKIM, DMARC), verifying sender reputation, assessing recipient-side filters, and optimizing email content. The guide also provides recommendations for maintaining a good sender reputation, such as warming up new domains, using dedicated sending domains, and implementing best practices for tracking and analytics. Additionally, the guide addresses common questions about DNS records, subdomains, and tracking pixels to help developers troubleshoot deliverability issues and improve their email marketing efforts.

WorkOS Vault bridges the gap for enterprise customers who want to implement Bring Your Own Key (BYOK) encryption, allowing them to securely manage their own keys without requiring your company to build a custom infrastructure. BYOK is essential for security-conscious organizations handling sensitive data, but implementing it can be complex and time-consuming. WorkOS Vault simplifies this process by providing a plug-and-play API and SDKs that abstract the complexity of BYOK, supporting seamless integration with multiple cloud providers' key management services (KMS), org-scoped key mapping, secure runtime access without storing keys, automatic key rotation and revocation, and built-in auditing and traceability. This solution helps enterprises meet compliance requirements while reducing operational risk and allowing you to focus on delivering great features to your users.

The first-ever MCP Night event took place on May 14, 2025, at the Exploratorium in San Francisco, bringing together hundreds of engineers, founders, researchers, and tinkerers to celebrate the Model Context Protocol (MCP). MCP enables large language models to connect with tools, APIs, data, and systems safely, securely, and in a standardized way. The event featured demos from top teams building exciting MCP servers and integrations, showcasing how MCP can be used for various applications such as network and edge automation, browser-based servers, personal assistants, customer contact finding, and workflows. The turnout exceeded expectations, with the main stage reaching capacity, and an overflow demo area was set up to accommodate the large number of attendees. After the talks concluded, the Exploratorium opened up for a museum-wide afterparty, featuring exhibits, music, and drinks, allowing attendees to continue conversations about AI, infrastructure, and auth in a casual setting.

SAML (Security Assertion Markup Language) is a legacy protocol that enables identity federation between an identity provider (IdP) and a service provider (SP). Implementing SAML from scratch can be challenging due to its complexity, ambiguity, and security pitfalls. The protocol's reliance on XML signatures, IdP quirks, and security vulnerabilities makes it prone to errors and maintenance issues. Common mistakes include trusting the wrong signature, skipping audience validation, improper XML parsing, and assuming metadata is static. To avoid these pitfalls, it is recommended to offload SAML implementation to a provider that can handle the complexity, such as WorkOS, which normalizes differences across identity providers and handles edge cases through a single, unified API.

Platforms like OpenAI Codex and Google Jules are taking a swing at distributed cognition for software teams by introducing agentic parallel coding, where entire development threads are offloaded to autonomous agents operating alongside GitHub repositories. These tools enable users to manage multiple repositories efficiently, with Codex providing an async engineering team interface and Jules offering a more polished and robust experience. The real story here is not about AI writing code, but about treating repositories as addressable environments for action, where tasks are launched, branches updated, and logs streamed in a way that's not currently possible with traditional coding tools.

Agno is an open-source framework that enables the construction of clean, composable, and Pythonic AI agents with tools, memory, and reasoning capabilities. It offers a declarative agent composition interface, where components are plug-and-play modules allowing for customization without vendor lock-in or framework ceremony. Built-in tools cover essentials like web search, financial data analysis, structured reasoning, code evaluation, and custom workflows, which can be mixed and matched or authored with minimal boilerplate. Agno also supports pluggable vector stores, short-term memory, long-term storage, and transparent reasoning, making it suitable for teams building agentic systems, startups shipping internal copilots, engineers experimenting with RAG, workflows, or multi-agent setups, and builders seeking full control over their agents' thinking, acting, and persisting. Getting started with Agno involves visiting the website and demo, accessing the GitHub repo, and installing the framework via pip.

Single-page applications (SPAs) offer fast navigation, rich interactivity, and efficient API communication, but this shift to the client has made security responsibilities more dangerous. SPAs hold tokens in browser storage, decide which UI elements to show based on locally stored roles, and manage route access with JavaScript, making security decisions vulnerable to attackers who can inspect, modify, or replay requests with little effort. Common threats include Cross-Site Scripting (XSS), token storage and theft, Broken Object-Level Authorization (BOLA), insecure routing and fake protection, and session and token expiry problems. To defend against these threats, developers should implement practical defenses such as sanitizing user input, storing tokens securely in HttpOnly cookies, enforcing object-level access control, protecting data access on the server, using schema validation libraries, rotating refresh tokens, and keeping third-party dependencies up to date. Additionally, outsourcing security concerns to a hosted login flow with proper cookie-based sessions can significantly reduce exposure to many of these risks.

To sync users from Microsoft Entra ID (formerly Azure AD) into an app using SCIM, Node.js, and WorkOS, follow these steps: Install the WorkOS Node SDK, set secrets with your WorkOS API key and client ID, configure the Entra ID connection by creating a new organization in the WorkOS dashboard and setting up a manual directory sync, and then connect Entra ID to WorkOS. Once connected, any time a directory-related resource is created, updated, or deleted in Entra ID, WorkOS will generate an event that your app can consume to stay in sync with the identity provider in real-time. You can use either polling the Events API for new events or setting up webhooks to receive them automatically in real time. To start consuming events, pick a starting place in the data set by using a cursor and then determine the event types you want to pull. For more information on how to sync users and groups to your app, including streaming events to Datadog, refer to the provided documentation and tutorials.

OAuth 2.1 is the latest version of the OAuth authorization framework, which aims to improve security and usability while simplifying the specification. It builds upon OAuth 2.0, consolidating key security improvements from multiple RFCs, deprecates legacy or unsafe flows, and provides clearer guidance for developers and implementers. The new specification introduces mandatory Proof Key for Code Exchange (PKCE) for all authorization code flows, removes implicit flow, formalizes how single-page apps can securely use refresh tokens, and emphasizes secure transport, scope minimization, and token storage to mitigate security risks associated with bearer tokens. The update also includes changes to redirect URIs, password grant, and implementation guidance for different app types. To ensure a smooth transition, OAuth 2.1 provides a migration checklist that outlines steps to take when updating existing implementations or building new ones.

Secure APIs are designed with security in mind from the start, avoiding common pitfalls such as lack of authentication and authorization controls, injection attacks, overexposure of data, improper rate limiting, and lack of logging and monitoring. Best practices include implementing TLS everywhere, enabling CORS safely, validating input and output, enforcing Content-Type headers strictly, using API keys and access tokens securely, adopting the principle of least privilege, keeping dependencies updated, versioning APIs, minimizing trust between services, and using security headers in responses. When consuming third-party APIs, vetting providers thoroughly, validating and sanitizing external responses, using scopes and permissions wisely, implementing timeouts and retries, sandboxing and testing integrations, monitoring for deprecations, monitoring usage and performance continuously, rate limiting requests, encrypting and storing responses securely, and building with intention are crucial to prevent security threats.

NIST, the National Institute of Standards and Technology, is a U.S. federal agency that has become a low-key powerhouse in cybersecurity, developing technical standards and guidelines for security, privacy, cryptography, and digital identity. Their frameworks are widely adopted and influence how we authenticate users, manage data privacy, design resilient infrastructure, and prepare for future threats like quantum computing and AI misuse. NIST's work is relevant to everyone from startups to federal agencies, especially in areas like authentication, data protection, and secure architecture. The agency's guidance is vendor-neutral, implementation-agnostic, and scalable, making it a trusted source for developers and security engineers designing systems or evaluating risk. NIST has led work on emerging technology areas such as post-quantum cryptography, Zero Trust architectures, and AI risk management, shaping how industries think about secure-by-default design. Their publications often provide the building blocks for actual implementation without being overly prescriptive. Despite some criticisms that their guidance can be dense or slow, NIST delivers stable, peer-reviewed guidance that's meant to scale from startups to federal agencies, making it a useful resource for informed decision-making in security.

RFC 9728 introduces standardized OAuth 2.0 resource metadata, allowing protected resources to describe themselves via metadata, just like authorization servers can. This enables clients to discover the capabilities of a protected resource through a standardized metadata endpoint, improving integrations, security, and discovery. The metadata is a simple JSON document published at a `.well-known` URI, containing information such as supported token types, required scopes, and token introspection URLs. The spec adds quality-of-life improvements for clients, including easier integrations, better security, and dynamic discovery. It was finally published after nearly nine years of development, with the use case catching up with the spec over time.

This is an in-depth guide on how to diagnose and debug common issues with SAML (Security Assertion Markup Language) assertion failures. The article provides a step-by-step approach, covering topics such as understanding the structure of a SAML assertion, identifying common causes of failure, and using tools like SAML-tracer to capture and decode SAML responses. It also includes troubleshooting scenarios, best practices for configuring IdP and SP settings, and advice on how to handle certificate management and metadata updates. The guide aims to empower developers with the knowledge and skills needed to successfully implement and troubleshoot SAML-based Single Sign-On (SSO) integrations.

Authentication in on-premises and hybrid environments is challenging due to layered constraints such as firewalls and NATs, air-gapped networks, custom configurations, legacy protocols, and fragmented infrastructure. Common pitfalls include assuming public endpoints are always available, hardcoding auth protocol assumptions, neglecting certificate and key rotation, forgetting about auditing and monitoring, and overcomplicating the deployment model. Best practices for on-prem and hybrid authentication involve designing for asynchronous and fault-tolerant workflows, prioritizing end-to-end security, offering deployment flexibility, investing in documentation and self-service tools, planning for lifecycle management, and embracing operational excellence and customer empathy. WorkOS can help by providing a unified API for every identity provider, secure network integration, built-in certificate and metadata management, comprehensive observability and diagnostics, and deployment flexibility.