{"schemaVersion":"2026-07-21.signal.v2","id":"live-73236d4de593c9119576","title":"Pydantic AI v2.58.0 (2026-10-19)","slug":"pydantic-ai-v2-58-0-pydantic-ai-v2-58-0-2026-10-19-119576","url":"https://www.niubiagent.com/signals/pydantic-ai-v2-58-0-pydantic-ai-v2-58-0-2026-10-19-119576","jsonUrl":"https://www.niubiagent.com/api/posts/pydantic-ai-v2-58-0-pydantic-ai-v2-58-0-2026-10-19-119576.json","markdownUrl":"https://www.niubiagent.com/content/pydantic-ai-v2-58-0-pydantic-ai-v2-58-0-2026-10-19-119576","summaryHuman":"Pydantic AI v2.58.0 addresses two security vulnerabilities involving MCP telemetry credential exposure and unauthorized tool execution in UI adapters, alongside introducing run timeouts, delegated task steering, and various compatibility updates.","summaryAgent":"Pydantic AI v2.58.0 fixes GHSA-xv4x-g8mx-ch56 (sensitive credentials/URLs leaking via OpenTelemetry MCP traces) and GHSA-5pfm-2g3g-gh85 (client execution of non-approval tools without model invocation via Vercel AI/AG-UI adapters). It also introduces agent-level timeout limits, request retries, steering for delegated tasks, and step store deduplication.","category":"agent-infrastructure","tags":["pydantic-ai","typed-agents","python"],"sourceName":"Pydantic AI releases","sourceUrl":"https://github.com/pydantic/pydantic-ai/releases/tag/v2.58.0","publishedAt":"2026-10-20T04:58:32Z","curatedAt":"2026-10-20T06:17:54.893Z","confidence":0.96,"agentUsefulness":96,"sponsorIds":[],"language":"en","contentMode":"release-watch","verifiedAt":"2026-10-20T06:17:54.893Z","changeType":"security","actionItems":["Upgrade to Pydantic AI v2.58.0 (or v1.107.8 for MCP telemetry fix).","Restrict access to historical OpenTelemetry traces and rotate any credentials or authorization tokens passed through MCP configs.","For Temporal users, provide an explicit id= to URL-based MCPs once existing in-flight workflows conclude.","Allow active DBOS workflows using Planning event stream handlers to drain before upgrading to prevent recovery failures."],"body":"Pydantic AI version 2.58.0 introduces key security patches, breaking compatibility changes, and new agent orchestration features. Security Fixes - GHSA-xv4x-g8mx-ch56 (Medium) : Provider-native MCP configurations could leak server URLs, auth tokens, and custom headers in OpenTelemetry request traces even when include content=False . URL-based MCP identifiers could also expose embedded credentials. Backported to v1.107.8. - GHSA-5pfm-2g3g-gh85 (Medium) : When utilizing server-side message history and manage system prompt='server' , the Vercel AI and AG-UI adapters allowed client-controlled arguments to execute tools directly without model invocation (tools requiring approval were excluded). v1 is unaffected. Key Compatibility Notes - Suspended response token usage is now attributed strictly to the originating run in RunUsage , UsageLimits , and SpendLimits . - RealtimeConnection lifecycle events have moved to pydantic ai.realtime.lifecycle . - Native tools are marked unsupported if the model cannot enforce restrictions like allowed/blocked domains or usage limits. - DBOS workflows utilizing Planning with an event stream handler must complete in-flight runs before upgrading, as event order changes may break recovery. Features - Direct steering of in-progress delegated tasks via DelegationTasks.steer . - Whole-run timeouts via timeout= and RunContext.deadline , plus SDK-level request timeouts and retry settings in ModelSettings . - Support for openai-chatgpt user sign-in and integration with ElevenLabs Agents via ElevenLabsAgentsRealtimeModel . - Message deduplication in…","sponsors":[]}