Moltbook is a Reddit clone where AI agents post, comment, and build religions. It claims 1.6 million users but fewer than 1% are active, it leaked 1.5 million API keys through a misconfigured database, and MIT Technology Review called it peak AI theater. Here is what it actually tells us about agent-to-agent communication.
The Model Context Protocol connects 17,000+ servers to AI agents. In January 2026, three CVEs in Anthropic’s own Git MCP server proved what security researchers had warned about: MCP is a ripe attack surface. This post covers the real vulnerabilities, how tool poisoning works, the OWASP MCP Top 10, and what you can do about it today.
OpenAI built its Frontier platform around a single idea: AI agents need onboarding, permissions, performance reviews, and clear reporting lines, just like the people sitting next to them. This framing shift from ‘deploy software’ to ‘manage a workforce’ has real consequences for how IT, HR, and compliance teams operate.
Goldman Sachs spent six months embedding Anthropic engineers inside its technology teams to build autonomous agents for trade reconciliation, client onboarding, and compliance. The agents run on Claude Opus 4.6 and manage operations across $2.5 trillion in assets under supervision. Here is what the deployment actually looks like, why Goldman picked Anthropic, and what it signals for the rest of finance.
On January 22, 2026, Singapore’s IMDA released the world’s first governance framework built specifically for AI agents. It covers four dimensions: risk bounding, human accountability, technical controls, and end-user responsibility. Unlike the EU AI Act, compliance is voluntary. Here is what the framework says, where it diverges from European regulation, and why DACH companies should care.
Anthropic launched a legal plugin for Claude Cowork that wiped $285 billion off software stocks in two days. The plugin automates contract review, NDA triage, and compliance workflows for in-house counsel. Here is what it actually does, who it threatens, and where the hype outpaces reality.
Radware’s ZombieAgent is the first zero-click indirect prompt injection that persists in an AI agent’s long-term memory, exfiltrating data character by character through pre-constructed URLs. All of it happens inside OpenAI’s cloud, invisible to enterprise security tools. This post breaks down the full attack chain, the failed defenses, and what memory poisoning means for every organization deploying AI agents.
Mercor’s APEX-Agents benchmark gave frontier AI models 480 real professional tasks from investment banking, management consulting, and corporate law. The top model, Claude Opus 4.6, scored 29.8%. GPT-5.2 hit 23%. No model cracked 34% in any single category. The benchmark directly challenges claims that AI agents are ready to replace knowledge workers.
Most ‘build an AI agent’ tutorials produce a chatbot with extra steps. This one builds a real agent that reasons, calls tools, and manages state. Covers LangGraph, OpenAI Agents SDK, and CrewAI with working code, then breaks down the five mistakes that kill every first agent.
Apple’s Xcode 26.3 adds native agentic coding with Anthropic’s Claude Agent SDK and OpenAI’s Codex, plus 20 MCP tools that expose builds, tests, previews, and documentation to any compatible agent. This is not autocomplete. Agents can now autonomously write code, build your project, run tests, render SwiftUI previews, and iterate on fixes without leaving the IDE.