Security

Claude Code v2.1.221 introduces a Focus view for cleaner chat interfaces, enhanced sandbox credential masking on Linux/WSL, improved permission checks and MCP server handling, and numerous bug fixes across VSCode, CLI, and desktop platforms. Key improvements include better cache efficiency for auto-mode decisions, refined plugin management, and enhanced session handling for background jobs. The release addresses security concerns, authentication issues, and user experience refinements across multiple platforms.

OpenClaw v2026.7.2-beta.7 introduces major stability and feature enhancements including state safety with quarantine stores and crash-recoverable snapshots, durable channel delivery across multiple platforms, session rewind/branching capabilities, interactive MCP Apps with dashboards, structured questions and approvals, meeting integration with Teams/Zoom/Google Meet, Wear OS companion app, and guided local inference setup. The release also adds Claude Opus 5 support, native app improvements across macOS/Linux, enhanced memory recall, scheduling with dynamic cadence, Fish Audio speech synthesis, and numerous security hardening fixes.

OpenClaw v2026.7.2-beta.6 introduces major enhancements for state safety, durable message delivery, session management, and multi-platform support. Key features include crash-recoverable data protection with quarantine stores, conversation branching and rewinding, interactive MCP Apps with dashboards, structured questions/approvals across platforms, meeting integration with Teams/Zoom/Google Meet, Wear OS companion app, and guided local inference setup. The release also adds Claude Opus 5 support, native apps for macOS/Linux, enhanced memory management, and improved scheduling with dynamic cadence.

OpenAI Agents Python v0.19.2 is a maintenance release focused on bug fixes and stability improvements across multiple subsystems. Key improvements include exposing original callables through wrappers, fixing guardrail result reporting, enhancing MCP credential redaction for security, improving memory session item counting, and resolving voice/audio dispatcher issues. The release also addresses sandbox environment preservation, modal path validation, realtime event payload handling, and tracing deadline management.

OpenAI Agents Python v0.19.1 is a maintenance release focused on bug fixes and stability improvements across sandbox path handling, audio diagnostics, WebSocket error handling, model execution, memory management, and tracing. Key fixes address streamed model input handling, guardrail task cancellation, SQLite session counting, and batch trace processing. Documentation updates include URL corrections, tracing integration links, and standardization of decorator usage in examples.

OpenClaw v2026.7.2-beta.5 introduces major reliability and safety improvements including crash-recoverable state management with quarantine stores, durable channel delivery across multiple platforms, session rewind/branching capabilities, interactive MCP Apps with dashboards, and structured question/approval workflows. The release adds support for new models (Claude Opus 5, Kimi K3, GPT Live), native apps for macOS/Linux/Wear OS, meeting integrations (Teams, Zoom, Google Meet), and guided local inference setup. Extensive security fixes address authorization, SQLite data safety, channel delivery persistence, and session lifecycle management.

OpenAI Agents Python SDK v0.19.0 introduces Programmatic Tool Calling, allowing supported models to generate JavaScript for coordinating tools with per-tool caller restrictions and structured outputs. The release adds a public decorators module with a shorter @tool alias, standardizes SDK configuration to accept both typed objects and dictionaries, hardens sensitive logging across all components, and improves compatibility with AnyLLM, LiteLLM, and Chat Completions while preserving session history on retries.

CrewAI 1.15.7a1 is a patch release addressing critical bug fixes and observability improvements. Key updates include fixing registry skills resolution through CrewAI+ client, recovering from GPT-5.6 tool errors, enabling tool calling on the Responses API, and preventing 404 errors for responses-only models. The release also adds skill usage event emission for better runtime observability and includes a security patch for a bedrock-agentcore dependency vulnerability.

CrewAI v1.15.7 release includes critical bug fixes for registry skills resolution, GPT-5.6 tool compatibility, and responses-only model routing. The update enhances observability by emitting skill usage events at runtime and addresses a security vulnerability in bedrock-agentcore. This patch ensures improved tool calling functionality across different API paths and model types.

Claude Code v2.1.217 introduces emoji shortcode autocomplete, improves transcript reliability with failure warnings, and fixes critical issues including memory leaks, Windows auto-update failures, and security vulnerabilities in session isolation. The release also enhances subagent management with concurrency caps and nesting controls, while addressing telemetry routing, screen reader accessibility, and various edge cases in session handling and transcript management.

Claude Code v2.1.216 is a maintenance release focused on stability and performance improvements. Key fixes address quadratic slowdown in long sessions, OAuth token handling, file isolation settings, and numerous UI/terminal rendering issues. The release also improves command validation, workflow handling, and background session management across multiple platforms including VSCode and web clients.

OpenClaw v2026.7.2-beta.3 introduces remote coding sessions on cloud workers, native automation capabilities for mobile and headless Linux, safer channel operations with improved Telegram and Signal handling, guided Control UI setup for model providers and channels, and enhanced gateway/session recovery. The release also includes new Linux packaging options (deb and AppImage), external gateway supervision mode, ClickClack integration improvements, and numerous UI/UX refinements across Control UI, including better chat layouts, keyboard shortcuts, and session management.

Claude Code v2.1.214 is a security and stability release that fixes critical permission-check bypasses across multiple shells (Bash, PowerShell, zsh), improves command parsing to prevent auto-approval of unsafe operations, and adds new features like the EndConversation tool and OpenTelemetry enhancements. The release addresses 50+ bugs including permission prompt timing issues, background session management, streaming failures, and platform-specific encoding problems on Windows.

OpenClaw v2026.7.2-beta.2 introduces remote coding sessions on cloud workers, native automation and mobile capabilities, safer channel operations, guided Control UI setup, improved gateway and session recovery, and enhanced Linux packaging. Key improvements include external gateway supervision, ClickClack integration, Skill Workshop enhancements, and numerous Control UI refinements for better user experience and reliability.

Onyx is an open-source chat UI (YC W24) that works with any LLM and provides enterprise-grade features like RAG, web search, MCP, and deep research. Built by the creators of Danswer, the team pivoted after realizing users wanted secure, self-hosted chat interfaces with world-class UX rather than just search. The platform supports RBAC, SSO, and on-prem hosting, with key innovations in context management and model-agnostic tool handling that enable reliable multi-model deployments at scale.

Forge is an open-source reliability layer for self-hosted LLM tool-calling that dramatically improves agentic task performance on consumer hardware. It uses domain-agnostic guardrails (retry nudges, step enforcement, error recovery, VRAM management) to boost an 8B model from 53% to 99% accuracy on multi-step workflows without model changes. The framework addresses the compounding failure problem in local models—90% per-step accuracy yields only 59% success on 5-step tasks—and peer-reviewed results show an 8B model with Forge (99.3%) outperforming Claude Sonnet without guardrails (87.2%).

Claude Code v2.1.211 is a maintenance release focused on bug fixes, security hardening, and performance improvements. Key updates include new subagent text forwarding options, fixes for permission preview character sanitization, auto mode hook decision handling, and credential store synchronization issues. The release also addresses file upload validation, terminal performance, background session management, and cross-platform compatibility issues across Claude in Chrome, VSCode, and various LLM backends (Vertex, Bedrock, Mantle, Foundry).

Claude Code v2.1.207 introduces auto mode availability without opt-in on Bedrock, Vertex AI, and Foundry, while fixing critical issues including terminal freezing during long streaming responses, security consent dialog bypasses, and various UI/workflow bugs. The release improves agent view functionality, Remote Control session handling, and Deep Research labeling, plus enhances security by rejecting shell-injection patterns in plugin hooks and preventing unintended settings persistence.

Claude Code v2.1.205 introduces critical bug fixes and improvements across session management, file handling, and auto mode functionality. Key enhancements include blocking tampering with session transcripts, fixing JSON schema validation issues, improving background agent state tracking, and optimizing memory usage during auto-updates. The release also strengthens security by preventing fabricated approvals and adds better diagnostics with the enhanced `/doctor` command.

OpenAI Agents Python v0.17.8 release introduces an invalid final output recovery handler and fixes critical issues with realtime session management, Pydantic validation, visualization, and sandbox execution. The update includes improvements to error handling, logging security, and data serialization compatibility. Documentation enhancements and dependency updates round out this maintenance release.