Prompt Injection Defense

crewAI v1.15.17 introduces declarative conversational flows with opt-in conversational mode, improves tool handling with AMP slug support and oversized message chunking, and fixes critical bugs including MCP server configuration, agent scope closure, tool error attribution, and OpenAI Responses API compatibility. The release enhances security with SSRF checks per redirect hop and resolves native tool call issues.

Claude Code v2.1.232 introduces major improvements to subagent forking, cross-session messaging, and security hardening. Key features include default subagent forking with prompt cache inheritance, @-mention syntax for direct session communication, GitLab marketplace support, and numerous security fixes for permission bypasses and credential handling. The release also enhances Remote Control stability, improves managed settings validation, and fixes various UI and performance issues.

OpenClaw v2026.6.34 is an extended-stable maintenance release focused on security hardening and reliability improvements without new features. Key enhancements include safer browser and network boundaries with sandboxed routes, more resilient agent and provider runs with improved session handling, stronger channel recovery mechanisms, safer operator diagnostics with credential protection, and more robust local runtime state management. The release includes 25 merged PRs addressing execution safety, delivery stability, and gateway reliability, with upcoming deprecations for Plugin SDK migration scheduled after July 24.

Claude Code v2.1.223 introduces owner wildcard marketplace controls, improves security by fixing permission bypass vulnerabilities and sandbox escapes, and enhances model discovery and session management. Key updates include stricter command obfuscation prevention, workflow sandbox isolation fixes, and better handling of context windows across Claude models. The release also refines the code review workflow and improves managed settings merging for enterprise deployments.

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 v2.1.207 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 v2.1.183 introduces significant safety improvements by blocking destructive commands (git reset, terraform destroy, etc.) unless explicitly requested, adds deprecation warnings for models, and includes the new `attribution.sessionUrl` setting to control commit/PR attribution. The release fixes multiple critical bugs including WebSearch failures in subagents, TUI corruption in Windows Terminal, and tmux teammate pane issues, while improving configuration management with `/config --help` and better toggle behavior.

Claude Fable 5 includes safety classifiers that block requests in cybersecurity, biology, and reasoning extraction areas. Anthropic provides server-side and client-side fallback mechanisms to automatically retry blocked requests with Claude Opus 4.8, along with new billing changes that credit fallback token costs. This guide covers classifier block detection, fallback implementation strategies, streaming behavior, and billing adjustments for API customers.

OpenClaw v2026.6.5-beta.2 introduces critical fixes for model reasoning leakage, MCP tool result handling, Anthropic extended-thinking recovery, and adds Parallel as a bundled web-search provider. The release improves durability of auth/plugin state via SQLite, enhances Matrix voice/thread support, and strengthens upgrade/service paths with safer migration and configuration handling. Multiple provider integrations (Google Vertex, WhatsApp, Mattermost) receive reliability improvements alongside comprehensive test isolation and CI/E2E enhancements.

Claude Code v2.1.149 introduces enhanced usage analytics with per-category breakdowns, improved keyboard navigation in diff views, and GFM task list rendering. The release includes enterprise features for MCP connectors and fixes numerous security vulnerabilities in PowerShell execution, sandbox permissions, and file system operations. Additionally, UI/UX improvements address transcript freezing, argument hints, and feedback reporting for better debugging in long sessions.

OpenClaw v2026.5.3-1 is a hotfix release for the npm core package addressing a security scanner issue. The fix prevents the install scanner from incorrectly blocking official bundled plugin packages when process.env access and API calls appear in different sections of the compiled bundle. The updated package is published on the beta dist-tag for testing before stable release.