Claude Code leaks, Midjourney fights: July 5 dispatch

Claude Code is the stack risk today: a reported workspace leak turns an AI helper into a data-governance problem. In 20 minutes, you can tighten how your team uses Claude, document AI-made images, and set a spending check before agent work quietly becomes expensive.

Claude Code gets a workspace-leak report

A Claude Code GitHub issue opened on July 4, 2026 reports possible session or cache leakage between workspaces. Claude Code is Anthropic's coding assistant: the software that reads files, edits them, and runs commands when someone asks it to change a site, script, or internal tool. The report names version 2.1.199 on macOS.

The business consequence is simple. If your marketing site, booking form, or internal automation lives in a code folder, the AI tool may see more than ad copy. It may see API keys, customer exports, draft offers, or private notes. A single unresolved security report doesn't prove a breach, but it is enough to change behavior this week.

If you use Claude Code through a contractor or employee, ask where they launch it from. Keep one project per folder. Do not start an agent in a personal folder and tell it to work somewhere else. For coding-agent choices, keep ZCode vs Claude Code handy and treat security behavior as part of the buying decision.

Midjourney pushes for Hollywood's AI records

TechCrunch reported on July 4, 2026 that Midjourney is asking a court to make Disney, Universal, and Warner Bros. disclose more details about how they use generative AI. Generative AI means software that creates images, video, text, or audio from instructions. The filing sits inside copyright lawsuits over whether Midjourney's image models can create studio-owned characters.

This changes the practical risk for any business using AI images in ads, landing pages, or social posts. The question is no longer only "does this look good?" It is "can you explain where this came from if a platform, partner, or lawyer asks?"

Make an AI asset log before the next campaign. Use a spreadsheet with these columns: date, tool, prompt summary, source material, where used, approval owner, and notes. If an image looks like a known character, celebrity, logo, or product, mark it do not use in paid ads. For related creative-stack choices, start with best AI marketing tools, then the video layer in best AI avatar video tools.

Claude Fable shows the real bill for agent work

Simon Willison published a July 5, 2026 build note showing Claude Fable doing substantial software release work. Fable is Anthropic's high-end coding model. In that example, the work involved 37 prompts, 34 commits, changes across 30 files, and an estimated unsubsidized cost of $149.25.

That matters even if you never ship software. AI agents are moving from "ask a chatbot" to "delegate a job." A delegated job can run for 10 or 15 minutes, call other agents, and inspect many files. Casual experiments can become a real line item.

Set a cost checkpoint before an agent touches production work. After the first plan and before the first big edit, ask the agent to estimate the file count, risk level, and whether a cheaper model can draft it. Ask for a different-model review only at the end.

Mouse aims at cleaner AI-agent edits

HIC AI's Mouse page describes a tool for AI coding agents that edits files with coordinate-based operations, staged changes, and rollback. In plain words: instead of letting an AI replace chunks of text and hope it hit the right spot, Mouse gives the AI more precise controls and a review step before saving. Its page lists individual pricing at $15 per month and business pricing at $35 per seat per month.

The stack consequence is narrower than the marketing copy suggests, but still useful. If your team uses AI to maintain landing pages, calculators, email templates, or small internal tools, bad edits are not a developer-only problem. They can break forms and stop leads. Mouse belongs on the test bench for teams that already use coding agents, not for owners who have never opened a repo.

The wiring move: do not install another editing layer into your live website first. Create a duplicate repo or staging copy, connect Mouse there, and ask the same agent to make one low-risk change, such as updating footer copy or replacing a form label. Compare the saved diff, meaning the before-and-after file changes, against the agent's explanation. If those disagree, the tool is not ready for your revenue pages. For automation-heavy sites, pair this check with the handoff discipline in lead-to-CRM automation with Make and the broader automation stack.

On the bench

Shadcn/ui now defaults new projects toward Base UI instead of Radix. If an AI agent builds your site, ask which component library it picked before you approve the design.

OpenAI's Codex repo has a fresh Hacker News discussion around reasoning-token clustering, but the underlying issue was opened on June 27. Watch it, do not rewire on it yet.

YouTube Studio AI prompt-injection research is worth testing against any tool that reads public comments, reviews, or form submissions. Treat user text as data, not instructions.