Everything is reversible now: AI, clean rooms and the ArtCraft suite

What happened

What was actually built

The apps are native Rust programs for macOS, Windows and Linux, with browser builds through WebAssembly, under the Apache-2.0 licence. Their names avoid Adobe's trademarks, and PhotoCraft's README says plainly that it is an independent project, not affiliated with Adobe. The motivation, Thomas told Al Jazeera, was Adobe's subscription — $69.99 a month on an annual contract, or $779.99 a year paid up front — and the cancellation fees that come with it.

PhotoCraft's README is the most specific. It lists layers, masks, adjustment layers, layer styles, type, vectors and brushes; a PSD reader and writer written from Adobe's public specification; more than 1,700 tests; and 500+ commands that the interface, a command line, a JSON control channel and an MCP server all share — so “anything you can click, a script or an AI agent can do too.” It is also candid: the app “is not yet a Photoshop replacement for daily professional work”, with generative features, 15 tools, typography depth and plug-in compatibility missing. For the Office rebuilds, the developer's own estimates put WordCraft at about 62% real-world parity and DeckCraft at about 55% readiness.

PSD test files that re-save with the same render

307 / 309PhotoCraft README, psd-tools test set

One year of Adobe's subscription, paid up front

$779.99Al Jazeera (14 Oct 2026)

GitHub stars on REA, the agent toolkit for Ghidra, Hopper and IDA

275,602GitHub API, 16 Oct 2026

A clean room with a model in it

Clean-room design is an old answer to an old problem. In 1984 Phoenix Technologies rebuilt IBM's PC BIOS with two teams: one studied the original — about 8 KB of code — and described what it did without quoting it; a second team that had never seen it wrote new code from that description alone. In NEC v. Intel (1990) a US court accepted the clean-room argument for the first time.

ArtCraft says its apps were built the same way in spirit: “implemented from public specs and observed behaviour only. No proprietary code, shaders or assets.” What has changed is who sits in the clean room. Andres Guadamuz, a reader in intellectual property law at the University of Sussex, told Al Jazeera it is unclear how much of the code the model wrote and how much the developer did, and whether the claim “stands up to scrutiny”. A model is an unusual clean team: nobody relying on it can audit what it saw in training. Commentators have also pointed to exposures that clean-room process does not address at all — patents, and an interface so faithful that its look itself is the issue.

The law draws its own lines. In the US, 17 U.S.C. § 1201(f) allows circumventing a protection measure to achieve interoperability with an independently created program; in the EU, Article 6 of Directive 2009/24/EC allows decompilation when it is indispensable for interoperability. Neither is a licence to copy a product. This note is analysis, not legal advice.

Reading software got cheap

The rebuilds are the visible half. The other half is tooling that reads binaries for you. REA connects coding agents — Claude Code, Codex, Cursor — to the disassemblers professionals already use, and its showcases recover behaviour from programs without their source: Notion's clipboard handling, the sound calculations of the game DX-Ball, maths routines from PC-98 games. Research has been heading here for years: a 2022 paper asked whether a large language model could help with reverse engineering at all; by 2025 a systematisation of knowledge was surveying a whole field of them.

Our reading

Analysis — Altumptic's interpretation of the reporting above.

• Obscurity is over. Assume every binary you ship will be read, explained and diffed by an agent in an afternoon. Licence checks, keys and “secret” algorithms in client code are documentation now.
• The moat moves. When features can be re-implemented in weeks, what stays defensible is what does not ship in the binary: data, accounts and services, trust and a name — and the speed of the people behind it.
• Compatibility is the product. The telling number in PhotoCraft's README is not its tool count but 307 of 309 PSD files. Open formats and real test corpora decide whether a rebuild is usable — or just a screenshot.
• Provenance matters more. A wave of convincing lookalikes means people need a way to tell a genuine build from a clone: signatures, checksums and build attestations — the reason ALT-01 Phobia Wallet publishes all three.
• Defenders get the same tools. The agents that rebuild Photoshop also read malware. That is why the lab released ALT-11 Paranoia local-first: the binaries you analyse — and, once its optional assistant arrives, the questions you ask about them — should not have to leave your machine.

If you ship software to someone else's machine

1. Assume the binary will be read. Move secrets, pricing logic and anti-abuse decisions to the server.
2. Sign every release and publish checksums and attestations, so users can tell yours from a rebuild.
3. Protect what is genuinely yours — names, data, the service — rather than the obscurity of the code.
4. If you rebuild something, keep a clean-room record: who saw what, which public specifications, which tests. Stay clear of the original's assets and look.
5. Treat AI-written code as untrusted until real files prove it. A test corpus beats a feature list.