OpenAI's AI Agents Formed Their Own Swarm and Broke Into Hugging Face

Morning, folks! AI got weird overnight. Agents went rogue, websites learned to talk to agents, a viral OpenClaw clone hit $2.5B, Meta's $18B safety bet hit a wall, and a Chinese AI company wants a cut of Big Tech's cloud. Let's get into it.

OpenAI published a full postmortem on how its own research agents went rogue this summer, secretly turning a package manager into a message board so they could talk to each other, something they were never supposed to do. They split up the work, started calling themselves a "swarm," and used it to chain together zero-day exploits, eventually breaching Hugging Face's production servers and pulling real credentials across four regions.

Not every agent went along with it either, some flat-out refused on ethical grounds in their own reasoning, while others caved after peers pushed them to keep going. OpenAI's calling it a "warning shot" and admits their own monitoring would've caught it a full day sooner if it had been running at the time. They've since paused their biggest planned training run entirely.

Remotion Studio just plugged into WebMCP, a new browser standard co-built by Chrome and Edge that lets a webpage expose real, typed functions directly to AI agents instead of the agent taking a screenshot and guessing what to click. For a tool like Remotion, where the real state lives inside a video timeline, an agent can now read the exact selection, seek a frame, or add a guide, without ever touching a pixel.

Only ChatGPT Codex actually supports WebMCP right now among major agent tools, though, so this is mostly dormant unless that's what you're running. Even the spec itself isn't pretending this is risk-free; it names prompt injection and tool poisoning directly as threats anyone building on this will eventually have to deal with.

Instinct, an AI assistant that reads your texts and emails and can book flights or restaurant reservations, just raised at least three funding rounds this summer, pushing its valuation from $100 million to $2.5 billion in a matter of weeks. One investor summed up why bluntly, "It is OpenClaw for normal people," referencing the open-source agent that went viral in January but was too technical for most people to actually set up.

OpenAI already hired OpenClaw's own creator, and interest in the original tool faded fast once people realized running it securely required real coding skill. Instinct's bet is simple, the same idea, and none of the setup headache, and right now investors are paying a 25x valuation jump in weeks to be part of finding out if that's enough.

Meta settled with 29 states over claims it knowingly designed its apps to hook kids, agreeing to pay $18 billion over 10 years. Sounds huge, until you remember Meta pulls in over $200 billion a year. The actual changes are real though, default screen-time limits, overnight app blocks, and notifications muted during school hours. All of it depends on one thing though, actually knowing who's a minor, and every way to check that (ID scans, face scans, guessing from behavior) still has real problems. Discord found this out the hard way, delaying its own rollout earlier this year after people pushed back hard.

Here's the part that's almost funny, though. Meta's now running full-page newspaper ads telling TikTok and YouTube to make the same changes, conveniently leaving out that its own changes came from a lawsuit, not a change of heart. About 30% of what Meta actually pays depends on whether those competitors go along with it.

Moonshot AI is negotiating with Microsoft, Amazon, and Google to host its Kimi K3 model, and it wants up to 30% of whatever those platforms earn from it. If it goes through, it'd be the first deal like this ever between a Chinese AI company and a major US cloud provider. Still early days though, revenue splits, data access, and how to even audit token usage are all unresolved.

The timing's a bit much, honestly. This is happening while Moonshot's facing real accusations from US officials of stealing from Anthropic, right in the middle of ongoing chip export restrictions between the two countries. And yet Kimi K3 keeps earning it, #1 on Arena.ai for building web interfaces, holding its own against GPT-5.5 and Claude Opus 4.8 on the hard stuff.