AI Can Now Clone Photoshop. Here's What Gets You Banned.

What happened

On 23 October a security researcher posted four words that 1.8 million people have now seen: “Banned for reverse engineering work.” He says OpenAI closed his account. About five hundred dollars of plan value and saved-up usage, gone. And he says all he did was reverse engineer software on his own machine.

The same week, the same kind of AI workflow was used to clone Photoshop and six other Adobe apps for free. So far, the person who cloned Adobe hasn’t faced any action. The person poking at his own machine lost his account.

That’s because there isn’t one rulebook but three: the software maker’s, the AI company’s and, at work, your employer’s. They don’t agree, and only one can take your account away in a single click. I’m an internal auditor, CIA and CISA. Reading fine print is my job, so I read all three.

The tool at the centre of this is REA, short for Reverse Engineer Anything. It plugs AI agents like Claude Code or OpenAI Codex into professional reverse engineering tools. Think of an app as a sealed restaurant kitchen: REA lets the AI taste every dish and work out the recipe. It went from under ten thousand GitHub stars to over a hundred thousand in six days. People pointed it at smart TVs and old games. Then someone pointed it at Adobe.

Brandon Thomas says Adobe charged him its cancellation fee more than once. That’s the same Adobe that agreed this year to a settlement worth a hundred and fifty million dollars over those fees, while denying wrongdoing. So he released seven free apps, built with Claude: PhotoCraft for Photoshop, FilmCraft for Premiere and the rest. His defence is one line in the README: “Built from public specs and observed behavior only.”

Where the control failed

Watching an app isn’t stealing, right? Copyright probably agrees. It protects code, not what an app does. Even India’s Copyright Act allows studying and testing a program to understand how it works.

But Adobe doesn’t need copyright. It has a contract. Section 17 of its terms of use bans reverse engineering and then defines it to include “monitoring or tracking the inputs and outputs in order to recreate that system.” Observed behaviour is the definition. And Futurism reports that ArtCraft’s own co-founder described the project as reverse engineering Adobe’s top seven tools with Claude.

In audit we’d call Section 17 a compensating control: a backup for when the main control fails. Adobe’s main protection was that reverse engineering was hard. AI broke that. The clause doesn’t stop cloning. It makes it a breach. And US courts have enforced clauses like this even where copyright would have allowed the copying.

Here’s the difference that matters to you. Adobe needs lawyers and a court to enforce its rulebook. The AI companies need one click. That’s where the researcher got caught. He says he never used OpenAI to find or exploit vulnerabilities, only for local reverse engineering, which he says tripped something called Daybreak. Daybreak is OpenAI’s programme built to let approved researchers do exactly this kind of security work. So the system built for researchers appears to have flagged a researcher.

Honest caveat: this is his account, and OpenAI hasn’t commented on his case. RuntimeWire notes that OpenAI’s own guidance says a safety-check notice alone isn’t a finding that you broke policy. But the next day someone asked on OpenAI’s developer forum for clear rules on reverse engineering devices you own, and Codex users have reported false “Cyber Abuse” flags since June.

The rulebook an auditor reads first

What about Claude? The same week, a widely viewed post claimed Anthropic had quietly killed REA, and users reported Claude stopping mid-answer on reverse engineering requests. What I found was refusals, not bans: not one confirmed Claude account ban over REA, and no confirmed change from Anthropic. Still, people are moving REA to local models to avoid bans, so the fear is real.

Anthropic’s updated usage policy, dated 12 November, says it does not prohibit security research on systems you own or operate, or with the owner’s permission. It also says hitting a block does not by itself mean you broke the policy. But the same policy says Anthropic can limit, suspend or terminate access for real violations.

So a refusal is a warning light. In audit terms, it’s a control doing its job. A ban is what can happen when you keep testing that control: rewording, jailbreaking, opening new chats to sneak around it. That’s how one block becomes a pattern.

Which leaves a question. If Adobe’s contract is so strong, why hasn’t Adobe used it? As of 25 October, three weeks in: no comment, no reported cease and desist, no lawsuit. My take is that suing a solo developer is free marketing for him, open source forks are already out there, and the market didn’t wait: HSBC and Wells Fargo both downgraded Adobe.

Notice what that means for you. Almost every paid app you use has a clause like Adobe’s. You clicked agree. The only question is whether anyone enforces it.

The rulebook that gets enforced fastest is your employer’s. If you use Adobe through your company’s plan, its terms treat the company as the customer, and Adobe can audit a business once a year with seven days’ notice. Picture someone at work pointing an AI agent at a licensed tool just to see how it exports a file. Nobody stole anything. But the company may have just breached a contract it signed, on a company laptop, with a company AI account. That’s a very different conversation from a refused prompt.

The test to run this week

If you work in IT risk or audit, check one thing on Monday: does your acceptable use policy even mention AI-driven reverse engineering? Most were written before this was a one-command install.

For everyone else, five rules to protect your account, your money and your job:

1. Stay in your lane. Your own code, your own device, open source software, and practice challenges built to be taken apart.
2. Never ask the AI to break a lock. License checks, DRM, paywalls and game cheats are a red line in REA’s own disclaimer and in Anthropic’s policy.
3. Treat a refusal as a stop sign, not a puzzle. Don’t hammer the same request in new words. If it’s a false flag, report it with the feedback button.
4. Don’t keep money you can’t afford to lose in one AI account. Prepaid credits and annual plans can vanish with a ban. The researcher learned that at five hundred dollars.
5. Before you install a free clone, ask who maintains it and who fixes its security holes. These apps are alpha by their own label.

If you do security work for a living, both companies run verification programmes for exactly this. Apply before you need it.

A researcher lost his account. Adobe hasn’t said a word. Both stories say the same thing. AI made taking software apart cheap, so the protection moved from code to fine print. The fine print we all click agree on without reading. Now you know where to look.