SSRF Vulnerability Found in AI SDK's OAuth Metadata Discovery Flow

·1 views

A security researcher discovered a server-side request forgery (SSRF) vulnerability in a popular AI SDK used to handle authentication for MCP servers, which allow AI applications to connect to external tools and data sources. The flaw existed in the SDK's OAuth metadata discovery process, where a developer-supplied server URL is fetched without any validation. While the codebase included a URL-validation function applied to credential submission endpoints, it was never called during the metadata discovery chain. A proof-of-concept test confirmed the SDK would connect to internal addresses, including 127.0.0.1, without restriction. The gap highlights how partial security controls can leave critical code paths exposed, especially in fast-moving AI tooling development.

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)

Prompt Fix, Not Code, Rescued an AI Debate Engine from Fake Disagreements

A developer building AdversarialDebate, an open-source multi-agent debate engine, discovered that despite solid architecture, the system was producing hollow debates where AI models simply acknowledged objections without genuinely engaging. The root cause was a poorly designed prompt that left models a low-effort escape route, resulting in an 89% 'theater rate' and near-zero concessions in early tests. A single prompt rewrite eliminated the problem, dropping theater rates to 0.2% across 411 full-corpus debates and generating over 8,800 genuine concessions. Version 0.2.1, released on August 28, 2026, further hardened the pipeline by adding row-count invariant assertions at five pipeline seams to prevent silent data loss. The update also introduced the first-ever recall measurements for the tool, reporting a 1.7–3.4% missed-issue rate, alongside 55 new unit tests.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why 'Unknown' Must Be a Payment State, Not Just a Handled Error

A technical analysis highlights a critical flaw in how payment systems handle HTTP timeouts: writing 'failed' to a database after a timeout is inaccurate, because the payment provider may have already processed the charge. When a request times out, there are three indistinguishable outcomes — the request never arrived, it arrived and succeeded, or it arrived and the response was lost — making any definitive status claim unreliable. The recommended fix is to treat 'unknown' as a distinct, durable state in the payment status enum rather than collapsing ambiguous outcomes into a binary success-or-failure model. The article also distinguishes between 'confirmed' (provider acknowledgement via API) and 'settled' (verified via bank statement), arguing that stopping at confirmed substitutes a promise for actual evidence. Properly encoding these states in the database schema is presented as essential for accurate reconciliation and preventing duplicate charges or erroneous refunds.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Next.js Auth Checks Inside Suspense Boundaries Can Leak UI to Logged-Out Users

A subtle timing flaw in Next.js App Router can expose protected dashboard UI to unauthenticated users when auth checks are placed inside Suspense boundaries. Because React begins streaming layout elements and skeleton fallbacks before an async session check resolves, a logged-out visitor may briefly see navigation structure and UI components before a redirect fires. The vulnerability does not expose actual user data, but it can reveal app structure, feature names, or navigation sections intended to remain private. The fix is straightforward: move the authentication check above and outside any Suspense boundary so the redirect decision is made before streaming begins. This ensures nothing from the protected route reaches the browser until authorization has been fully resolved.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Route Pingvera Monitoring Alerts to a Help Desk Using Webhooks

A best-practice guide outlines how to integrate Pingvera monitoring events with help desk systems using secure HTTPS webhooks. The approach recommends creating a single incident ticket per confirmed failure, updating it as conditions change, and closing it only after verified recovery — avoiding duplicate tickets from retried probes. Key implementation steps include authenticating and persisting raw events before acknowledgment, processing them asynchronously, deduplicating by delivery ID, and mapping incidents to tickets through a stable internal schema. Security measures such as HMAC signature verification, replay detection, constant-time comparison, and secret management are strongly emphasized. The guide also stresses the need to monitor the integration pipeline itself, maintain a dead-letter queue, and keep a secondary alerting route for critical failures.

0 comments Read more at DEV Community