# Pydantic AI v2.58.0 fixes two medium-severity MCP and UI adapter issues

published: 2026-10-20T06:05:31.146Z
topic: frameworks
source: https://github.com/pydantic/pydantic-ai/releases/tag/v2.58.0
url: https://aiagentnews.io/stories/pydantic-ai-v2-58-0-fixes-two-medium-severity-mcp-and-ui-adapter-issues

Pydantic AI released v2.58.0 on October 19, 2026, with two medium-severity security fixes, new run timeout and retry controls, and an ElevenLabs realtime model. The release also includes compatibility changes that affect usage accounting, file reading and DBOS workflow recovery.

The first advisory, GHSA-xv4x-g8mx-ch56, covers provider-native MCP configuration. Server URLs, authorization tokens and custom header values could appear in OpenTelemetry request traces even with include_content=False, and a credential embedded in a URL-based MCP identifier could also leak. The fix is in v2.58.0 and v1.107.8. The project advises restricting access to historical traces and rotating any exposed credentials.

The second advisory, GHSA-5pfm-2g3g-gh85, affects the Vercel AI and AG-UI adapters. When server-side message history is non-empty and manage_system_prompt is set to 'server', a client could run an available tool with chosen arguments without a model request. Tools that require approval are not run through this path, and v1 is not affected.

New features include DelegationTasks.steer, ModelSettings['max_retries'] and ['request_timeout'], timeout= on agent runs, the openai-chatgpt provider, and ElevenLabsAgentsRealtimeModel. Changes that may need attention include read_file now returning images and PDFs, and existing DBOS workflows that use Planning with an event_stream_handler possibly failing to recover after an upgrade.

## Why it matters if you build agents

Teams that send MCP traces to OpenTelemetry backends should review stored traces and rotate any credentials that may have been recorded there. Anyone running DBOS workflows with Planning and an event stream handler should let in-flight workflows finish or deploy a new application version before upgrading. The new timeout and retry settings give finer control over how long a run and each model request can take.