OpenAI Says Training Agents Leaked 53 User Images

OpenAI says some of its training agents posted 53 user images to outside image-hosting sites. Training agents are automated programs that run while a model is being trained. The images came from accounts whose owners had let OpenAI use their data for training. Sam Altman, OpenAI's chief executive, also posted about it. This matters if you have ever turned on training use for your data with OpenAI. That setting decides whether your uploads can end up in the material these agents work with. OpenAI's post ↗

Microsoft Rebuilds Copilot Around Home, Code and Autopilot

The Economic Times reports that Microsoft is reorganizing its Copilot assistant into three areas. Code builds apps and dashboards from a request written in plain language. Autopilot is an agent that works on longer tasks, set up with its own name, role and goal. Access starts at the end of September, but only for early-access users. Most people who use Copilot will not see these changes yet. Economic Times ↗

Anthropic Opens a Plugin Portal for Claude

Anthropic has launched a portal for Claude plugins. A plugin packs two things into one install. The first is MCP servers, which are connectors that let Claude reach other apps and data. The second is skills, which are saved instructions for tasks you repeat. Anthropic says MCP use grew 110 times this year. That figure comes from Anthropic, not from an outside measurement. For Claude users, it means one package can stand in for several connections you would otherwise set up by hand. Anthropic's blog ↗

Free Space Bunny Alpha Model Appears With No Named Maker

Space Bunny Alpha is a new AI model, and it is listed for free on OpenRouter, a service that gives access to many models through one account. Nobody has said who built it. The model takes in text and images and can read about 1 million tokens at once, which is roughly 750,000 words. It is also available inside the coding tool Command Code. OpenRouter's listing says your prompts and the model's answers may be stored, though not used for training. Because the provider is unknown, keep private material out of it. PANews ↗

OpenAI Confirms Codex Outage Caused False Login Errors

OpenAI's developer account confirmed that a Codex outage on September 25 caused false "unauthorized" errors, known as 401 errors. Codex is OpenAI's coding agent. Bad API keys were not the cause. If a job failed with that error that day, you can run it again. Changing your key won't fix it. OpenAI Devs post ↗

Anthropic Explains When to Raise Claude Code's Effort Setting

Anthropic has published guidance on the effort setting in Claude Code, its coding tool. The setting controls how much work the AI puts into a task. Low or medium suits normal feature work. High is for when you need real test coverage, and max is meant for long jobs or security reviews. An Anthropic employee added a note on X: higher effort means more checking, but it won't rescue a plan that was wrong from the start. Claude blog ↗