Author: IBL News

IBL News | New York

With over 200,000 GitHub stars in 84 days, OpenClaw has become the fastest-growing software repo in history. It heralds an AI agent explosion. Just as LLM agents were a new layer on top of LLMs, claws are a new layer on top of LLM agents, elevating orchestration, scheduling, context, tool calls, and persistence to the next level.

This software, created to run 24/7 fully autonomous agents and developed by the Austrian programmer Peter Steinberger, now working for OpenAI, continues to grow rapidly.

An increasing number of developers are running AI agents on Mac minis, old gaming PCs, post-viral TikToks, managing their entire inboxes, and even controlling smart homes.

The security, a huge concern from the beginning, has been dramatically improved for production with IronClaw, written in Rust by security researchers. IronClaw has been created for high-security environments handling sensitive data, where prompt injection and data exfiltration are real threats. It promises enterprise-grade security in an open-source agent.

The security architecture has five layers, each a hard boundary:

• Layer 1: Network. TLS 1.3 encryption, SSRF protection, and rate limiting per tool.
• Layer 2: Request filtering. Endpoint allowlisting (HTTP requests restricted to explicitly approved hosts/paths), prompt injection pattern detection, and content sanitization.
• Layer 3: Credential management. Secrets encrypted with AES-256-GCM, injected at host boundaries. Tools never see raw credentials. 22 regex patterns with Aho-Corasick optimization scan all requests and responses for credential leaks in real-time.
• Layer 4: WASM sandbox. Untrusted tools run in isolated WebAssembly containers with capability-based permissions. No ambient access to the system.
• Layer 5: Docker isolation. Intensive tasks run in Docker containers with per-job resource limits (CPU, memory, execution time).

Essentially, OpenClaw is an LLM that can do things, with the Claude Agent SDK managing the entire loop (reason, act, observe, repeat) until the task is done, with platform-specific messaging output (Telegram, WhatsApp, Discord, etc.), an agent loop can chain together complex multi-step tasks (read a file, find a bug, fix it, run tests, check if they pass, and report back, with a hard cap of 20 iterations), and a Memory agent which has its identity personality, and hard boundaries (“never execute financial transactions”), and persists knowledge across conversations, using human-readable, and human-editable markdown files.

Users define new capabilities as SKILL.md files, markdown documents with instructions that the agent understands and executes the workflow.

This architecture (agent loop + memory + skills) is the shared DNA, and it’s MCP-first. Web search, file operations, image generation, and code execution are all external MCP tool servers that the agent connects to at startup. Adding a new capability means plugging in a new MCP server rather than modifying the core codebase.

The personality system uses seven Markdown files (AGENTS.md, SOUL.md, USER.md, TOOLS.md, IDENTITY.md, HEARTBEAT.md, MEMORY.md) that define the agent’s behavior. Changing your agent’s personality means editing a text file, not writing code.

OpenClaw has several models (Claude, GPT, DeepSeek, Ollama, Mistral, etc.) and over 11 messaging platforms (WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Matrix, Teams, Google Chat, Zalo, WebChat).

The ecosystem presents over 5,700 skills on ClawHub, with agents able to do almost anything: manage Gmail and Calendar (Gog skill), summarize web pages and PDFs (Summarize skill), automate GitHub workflows, generate images, edit PDFs, control smart home devices, and track crypto portfolios. Skills install with a single command and extend the agent without touching core code.

https://t.co/NRR3PkF5is

— Misbah Syed (@MisbahSy) February 22, 2026

https://t.co/5gwA1IWRuL

— Nate.Google (@Nate_Google_) February 23, 2026