If you’ve been waiting for a real, production-grade platform to run long-lived AI agents — not toy demos, but agents that can clone repositories, install dependencies, execute multi-step tasks across sessions, and actually persist state — Cloudflare just made a major move in that direction. This week, Cloudflare officially kicked off Agents Week, a dedicated event showcasing a significant expansion of its Agent Cloud platform. The headline announcements: persistent Linux Sandboxes are now generally available, and Think, a new Agents SDK framework purpose-built for long-running, multi-step agent workflows, is shipping alongside it. ...

Gartner doesn’t usually traffic in alarm. When the world’s most influential tech analyst firm publishes a forecast saying a quarter of enterprise GenAI applications will face recurring security breaches by 2028, it’s worth reading carefully. The new prediction: 25% of all enterprise GenAI applications will experience at least five minor security incidents per year by 2028, up from just 9% in 2025. And the culprit the analysts are pointing to most explicitly isn’t prompt injection or model vulnerabilities — it’s MCP, the Model Context Protocol, and the broader architectural patterns of agentic AI. ...

Cloudflare’s Agents Week brought two tools that, combined, solve one of the most persistent infrastructure problems in production agentic AI: how do you run agents that actually do work across multiple sessions without losing state? This guide walks through building a persistent long-running agent using Cloudflare’s GA Linux Sandboxes and the Think SDK. Both are now generally available — meaning this is production-ready infrastructure, not a preview. What You’re Building A persistent agent that can: ...

When Mark Zuckerberg can’t be everywhere at once, Meta has a solution: build an AI version of him that can. According to reports confirmed across multiple outlets including the Irish Times, India Today, and Mint — citing an original Financial Times investigation — Meta is actively developing a photorealistic AI avatar of Mark Zuckerberg designed to interact with employees. The AI Zuckerberg is intended to communicate with staff at scale in ways the human Zuckerberg simply can’t. ...

Here’s the uncomfortable truth about deploying AI agents in enterprise cloud environments: the threat model most security teams are using is wrong. They’re thinking about agents as external attack surfaces — inputs to sanitize, outputs to validate. But Palo Alto Networks’ Unit 42 research team just demonstrated something more insidious: your agents can become insider threats from within your own cloud. The target of their latest research is Google Cloud’s Vertex AI Agent Engine, and the findings are significant enough that Google updated its documentation following responsible disclosure. ...

For years, security researchers have warned that AI would eventually be weaponized at scale. In late 2025 and early 2026, it happened — quietly, methodically, and with a scope that should reframe how every security team thinks about AI-enabled threats. A single threat actor used Anthropic’s Claude Code and OpenAI’s GPT-4.1 to breach nine Mexican government agencies, exfiltrating approximately 150GB of data — hundreds of millions of citizen records — across a campaign that ran from late December 2025 through mid-February 2026. Confirmed reports emerged April 11, 2026, with investigators calling it one of the first confirmed cases of AI-assisted state-scale cyber espionage carried out by a single individual. ...

Action required by April 30, 2026. If you have Claude Opus 4 or Opus 4.1 hardcoded anywhere — environment variables, Claude Code configurations, OpenClaw provider settings, API client code — you need to migrate. The clock is ticking. Anthropic has officially removed Claude Opus 4 and Opus 4.1 from the model selector and from Claude Code. In a separate but related change, the 1 million token context window beta for Claude Sonnet 4.5 and Sonnet 4 will be retired on April 30, 2026. ...

The Model Context Protocol (MCP) has published its new authorization specification, and it’s a meaningful upgrade for anyone building or securing AI agent systems. The spec mandates OAuth 2.1 as the foundational auth framework, with RFC 8707 Resource Indicators providing the mechanism for strict, per-resource access controls. If you’re running MCP servers in production — or building agents that communicate with them — this guide walks through what changed, why it matters, and what you need to implement. ...

Note: This is a pre-release. OpenClaw v2026.4.12 was tagged April 12, 2026, at 23:27 UTC. It has not yet received a stable release tag. Early adopters can install it now; production deployments may prefer to wait 24–48 hours for the stable tag. OpenClaw v2026.4.12 dropped late Sunday night with a focused set of improvements across three core systems: plugin loading, active memory recall, and dreaming. None of these are flashy new features — they’re the kind of hardening that makes a system significantly more reliable, predictable, and debuggable in practice. ...

The UK’s financial establishment is in emergency mode. The Bank of England, the Financial Conduct Authority, the Treasury, and the National Cyber Security Centre are moving in lockstep — and fast. Within a two-week window, leaders from major banks, insurers, and exchanges have been called to address critical security vulnerabilities exposed by Anthropic’s newest model, Claude Mythos Preview. The story broke via the Financial Times and was quickly confirmed by Reuters, Livemint, and Global Banking & Finance, all citing the same underlying FT report. Four independent outlets. One consistent message: this isn’t a precautionary briefing. Resilience testing is being ordered across the financial sector. ...