Refacto Agents

Industry story

Update: OpenAI Agents Python SDK v0.19.2 patches MCP security and memory bugs

agent-framework guardrails mcp reliability tool-use

What's new since we last covered this: MCP security vulnerabilities patched in SDK release.

Analysis

Showing the shorter version.

OpenAI released v0.19.2 of its Python Agents SDK, a framework for building agentic applications, with a batch of bug fixes targeting MCP (Model Context Protocol, the emerging standard for connecting agents to external tools and data sources), memory session management, voice, and tracing. Notably, multiple fixes address credential and error redaction in MCP transport layers — preventing URL credentials and internal error details from leaking into traces and tool metadata — and a fix ensures failed MCP servers are cleaned up before reconnection attempts. Memory session fixes correct item-counting logic across SQLAlchemy, MongoDB, Redis, and Dapr backends, and input guardrail results are now correctly reported even when a tripwire aborts an agent run.

Comments