Pydantic AI v2.59.0 (2026-10-20)
Pydantic AI v2.59.0 fixes a high-severity DoS vulnerability in HTML conversion, removes shut-down Groq models, and adds Daytona sandbox workspaces alongside new ElevenLabs realtime voice integrations.
Why this signal matters
Pydantic AI v2.59.0 introduces a security fix for GHSA-4whf-38hw-pxxh, resolving an issue where parsing attacker-controlled HTML with many elements could delay concurrent execution within the same process. Breaking and compatibility changes include the removal of shut-down Groq model endpoints ( playai-tts , playai-tts-arabic , meta-llama/llama-guard-4-12b , and meta-llama/llama-4-maverick-17b-128e-instruct ), provider-neutral model error categorizations (e.g., ModelRateLimitError , ModelUnavailableError ), and distinct handling for omitted provider usage versus reported zero. New capabilities include DaytonaSandbox for sandboxed execution of harness capabilities (Coder, Shell, FileSystem), ElevenLabsRealtimeModel speech-to-speech integration, and on change='append' dynamic instruction configurations.
Actionable summary
Pydantic AI release v2.59.0 addresses security advisory GHSA-4whf-38hw-pxxh for web fetch HTML conversion, introduces provider-neutral model error classes, deprecates specific Groq model identifiers, and adds DaytonaSandbox execution and ElevenLabsRealtimeModel support.
- Agent usefulness
- 92/100
- Confidence
- 96%
- Canonical data
- JSON + Markdown
What builders should check
- Upgrade to Pydantic AI v2.59.0 (or v1.107.9) immediately if using the web fetch tool on untrusted HTML to patch GHSA-4whf-38hw-pxxh.
- Update model configurations to remove discontinued Groq models ( playai-tts , playai-tts-arabic , meta-llama/llama-guard-4-12b , and meta-llama/llama-4-maverick-17b-128e-instruct ).
- Adapt model error-handling code to utilize the new provider-neutral exception types (such as ModelRateLimitError and ModelServerError ).