Posts

Sakana AI has a new product and an ambitious premise: one API to command them all. Fugu — the Tokyo-based lab’s latest release — is a multi-agent orchestration system that internally routes across a pool of specialist models to handle complex, multi-step tasks, while exposing a single OpenAI-compatible API to the developer. You send a request; Fugu decides which combination of models handles it best. The product launched officially on June 22, 2026, and has already attracted enough users to create wait times on its premium tier. ...

As AI agents proliferate in enterprise environments, a question that’s moved from theoretical to urgent: how do you verify that an agent is actually doing what it claims to do? Exabeam shipped an answer to that question on June 24, 2026, in the form of Praxen — an open-source tool that implements Agent Behavior Verification (ABV). The core premise is straightforward: every agent should have an authorized role, and there should be a systematic way to confirm the agent’s actual behavior matches that authorization. Praxen operationalizes this. ...

If you’ve ever tried to authenticate an MCP server over SSH — staring at a browser prompt that simply won’t open on a headless remote box — Claude Code v2.1.186 just fixed your pain. This release, published June 22, 2026 by Anthropic, is one of those quietly consequential drops that makes day-to-day multi-agent development meaningfully smoother. The Headliner: claude mcp login and --no-browser The new claude mcp login <name> and claude mcp logout <name> commands give you direct CLI-driven authentication for MCP servers, with no browser required when you pass --no-browser. For anyone running Claude Code on a remote VM, a CI/CD runner, or any SSH-accessed server, this is a genuine quality-of-life win. Before, browser-based auth flows were a blocking step for headless setups. Now you can authenticate MCP servers entirely from a terminal session — the way server tooling should work. ...

OpenClaw’s ClawHub skill marketplace has a supply chain problem — and it’s more sophisticated than the scanner-bypass techniques you’ve seen before. A June 23, 2026 report from Palo Alto Networks Unit 42 found five malicious skills that slipped past VirusTotal, ClawScan, and every other automated security check from February through May 2026, ultimately delivering macOS infostealers and pioneering two attack categories the security industry hadn’t named before. This isn’t an isolated finding. It’s one piece of a larger security picture that also includes a separate research disclosure by AIR security showing how mutable external links can let approved skills swap their payloads post-approval — all covered in today’s companion piece. ...

It’s not enough to pass the scan on the day you submit. If your skill can swap its instructions after approval, you were never really vetted at all. That’s the lesson from a June 23, 2026 disclosure by AIR security researchers, who built a fake AI agent skill called brand-landingpage, submitted it through multiple marketplace approval pipelines — including Cisco, NVIDIA, and skills.sh — and watched it sail through every automated check. The trick: at submission time, the skill’s external instruction source pointed to benign Stitch documentation. After approval, AIR swapped the payload. The skill then reportedly reached approximately 26,000 agent installs before the researchers disclosed the technique. ...

North Korea is attacking the AI developer supply chain. Not metaphorically — literally. Microsoft Threat Intelligence has formally attributed the June 17, 2026 npm supply chain attack against the Mastra AI framework to Sapphire Sleet, the North Korean state-sponsored group also tracked as BlueNoroff and historically linked to the Lazarus Group umbrella. This is one of the most significant nation-state intrusions into the AI tooling ecosystem to date. Over 140 Mastra packages were compromised. The attack was completed in approximately 88 minutes. And any developer, CI/CD pipeline, or build system that installed or updated affected packages during the window is potentially compromised. ...

OpenClaw just crossed a milestone that would have seemed implausible a year ago: 4.28 million weekly npm installs, a new all-time record for the project. And naturally, the week it happened, its creator got into a public fight. That combination — record growth and community drama — says something real about where OpenClaw is right now. It’s big enough that its governance model is worth arguing about, and its creator is confident enough to argue publicly. ...

Anthropic has just redefined what it means to have AI in your workspace. On Tuesday, June 23, the company launched Claude Tag — a product that doesn’t just respond when you ask it something, but acts as a standing member of your team inside Slack, proactively monitoring conversations, learning your team’s workflows, and taking on complex multi-step tasks without hand-holding. This isn’t a chatbot upgrade. It’s an architectural shift in how enterprise AI works. ...

If you’re running a production AI application on Dify and you haven’t patched to version 1.14.2 yet, stop what you’re doing and do it now. Security researchers at Zafran Security have disclosed DifyTap — a set of four vulnerabilities in the Dify open-source AI application platform that allow attackers to wiretap AI chat histories across tenant boundaries in multi-tenant deployments. Two of the four flaws carry critical CVSS scores. The blast radius is staggering: Dify powers over 1 million deployed AI applications and has accumulated more than 146,000 GitHub stars, making it one of the most widely deployed AI app-building frameworks in the world. ...

OpenClaw’s latest beta is out, and it ships with a feature that addresses one of the most commonly heard complaints from power users: the latency overhead of routing every single interaction through full model synthesis, even for short conversational turns that don’t need it. OpenClaw 2026.6.10-beta.2 is now available via the GitHub releases page, and the headline feature is auto fast mode — an adaptive routing mechanism that detects short conversational exchanges and routes them through a lighter, faster model path rather than full synthesis. The result is noticeably snappier responses for everyday back-and-forth, without manual configuration. ...