AI reverse engineering: software's Robin Hood moment?

The new frontier of reverse engineering is not a machine that swallows a proprietary programme and returns its open-source twin a few minutes later. It is something less theatrical and, for that reason, more destabilising: AI agents that can query decompilers, trace calls, observe an application's behaviour and turn those clues into a new implementation.

REA, Reverse Engineer Anything, has become the symbol of this shift. Its documentation describes understanding and rebuilding individual features with inspectable evidence, not automatically copying an entire suite. In the same climate, clean-room projects such as PhotoCraft have appeared. PhotoCraft says it reproduces image-editing workflows from public specifications and observed behaviour without using proprietary code or assets.

Decompiling, observing and reimplementing are not synonyms, and clean-room is not a magic phrase. Yet the social signal is unmistakable: many see them as digital Robin Hoods confronting SaaS castles. This attraction comes from an economic model that many users regard as expensive, opaque and increasingly beyond negotiation.

Key takeaways

• AI is lowering the cost of functional reconstruction, but it has not eliminated the cost of quality, security and maintenance.
• Support for open-source clones is also a reaction to recurring prices, format dependency and roadmaps decided unilaterally.
• Democratic access and infringement can coexist in the same project: a social purpose does not automatically make the method lawful.

The indictment: when a subscription becomes a toll

Price is not a footnote when software is the tool used to study or earn a living. Figures vary by country, promotion and plan, but the orders of magnitude are real (prices checked on 20 October 2026). Adobe's official page on individual plan changes listed Creative Cloud Pro at EUR 928.76 a year when prepaid, and an annual renewal price of EUR 490 for students and teachers. Autodesk's Media & Entertainment Collection was listed at EUR 2,935 a year, compared with combinations of standalone products costing more than EUR 4,600 or EUR 6,600.

Not every product costs thousands: many companies provide education plans, reduced editions or free trials. But a creative or technical workflow rarely depends on one product. Add graphics, editing, modelling, collaboration and storage, and the right to work with industry-standard tools can become a permanent toll.

Paying does not mean governing

The issue is asymmetric power: a subscription grants access to the product, not a vote on its roadmap or a guarantee that a niche bug or interoperability request will be addressed on the user's timetable. In an open project, at least in principle, a community can inspect, fix, fork or fund a priority. In closed software, the user waits outside the drawbridge.

GIMP and LibreOffice must chase formats, habits, plugins and training material accumulated by commercial leaders. Blender shows that open software can escape a niche and how much governance, funding and continuity success requires. In the 2024 Global Spotlight Insights Report, half of respondents who contributed to open-source projects said that most of their work was unpaid; 63% cited lack of time and 34% lack of funding or compensation. The gap consists of paid hours, testing, support, documentation and distribution.

Students and young professionals at the locked gate

For someone starting out, a licence does not compete with a corporate budget but with rent, tuition and hardware. Education schemes ease the problem without removing it: they require eligibility, carry specific terms and may end just as a student tries to turn a skill into paid work. People do not necessarily learn the best tool for understanding a craft; they learn the one required by job adverts, courses and the files exchanged inside companies.

This is where the piracy taboo appears. When lawful access feels out of reach and product knowledge is presented as a professional prerequisite, a crack becomes, for some, an educational shortcut. It is an unlawful shortcut that may require disabling protections, running key generators and trusting anonymous distributors.

This does not prove that every crack contains malware. In September 2026, Microsoft Security documented a campaign using counterfeit download sites to distribute malicious installers, with victims including organisations in education and technology. A genuine open alternative that is signed, inspectable and maintained offers more than savings: it moves users away from a black market where the hidden price may be access to their computer, projects and credentials.

AI-built clones: revolution or prototypes mistaken for products?

Agents can generate structure, tests, parsers, interfaces and documentation, then correct themselves through execution. The arXiv preprint MirrorCode: AI can rebuild entire programs from behavior alone measured this ability across 25 programmes: the strongest model scored 56% on the benchmark and reimplemented a bioinformatics toolkit of roughly 16,000 lines. The same research, however, records 19 days and about USD 2,600 of inference for one attempt at a large task. That is a radical reduction against some estimates of human effort, not zero cost.

In a few hours, an agent can produce an impressive demo or reconstruct a bounded feature. A professional alternative needs compatibility, predictable performance, accessibility, security and accountability for updates. PhotoCraft reports more than 1,700 tests and a Rust architecture. Those are signs of seriousness, not proof of complete parity with Photoshop.

Democratisation does not compile itself

Interoperable formats, local tools, lower prices and inspectable code can return power to schools, freelancers and small firms. A community can adapt an application for languages, assistive devices or systems neglected by the main market. But if a project depends on one developer, cloud credits or outputs that few people can review, the new Robin Hood risks handing out keys that nobody knows how to maintain.

AI does not remove open-source work. It shifts the centre of gravity from typing to specification, from the first implementation to review, and from scarcity of code to scarcity of trust. The Lehora Studio articles section can continue that discussion around the cost, maintenance and governance of digital products.

Reverse engineering and the ethical dilemma: a good cause is not a licence

Directive 2009/24/EC allows a lawful user, under defined conditions, to observe and study a programme, and permits decompilation when it is indispensable to the interoperability of an independently created programme. The information may not be used to develop or market a programme substantially similar in its expression, or for other acts that infringe copyright.

In SAS Institute v World Programming, the Court of Justice of the European Union held that functionality, a programming language and data-file formats are not protected by copyright as the expression of the programme. Code, assets, manuals, trade marks, trade secrets, patents and technical protection measures may trigger different disputes. An EULA matters too, although a contract clause cannot automatically erase mandatory exceptions provided by the applicable law.

REA itself says it is intended for lawful research, education, interoperability, preservation and the analysis of software that the user owns or is authorised to examine. As a risk control, a credible clean-room process separates behavioural study from implementation, keeps specifications and provenance records, excludes original code and assets, and subjects the result to legal review.

Questions Robin Hood's cloak cannot hide

Is it morally acceptable to clone somebody else's work when the original market is perceived as predatory? Does an unfair price turn infringement into justice, or merely explain why that infringement wins public sympathy? Where is the boundary between reproducing a function required for compatibility and appropriating the creative decisions that make a product recognisable?

Is it ethical to use agents to rewrite a free project in order to strip away attribution or copyleft obligations? Ethics cannot depend on whether the target is a popular or unpopular multinational. It must examine method, proportionality, authorisation, harm and public interest.

If code becomes abundant, what will companies sell?

It is far from certain that almost all software will become open source. But many reproducible features may lose value as competitive moats. Companies will then have to sell what a clone cannot obtain from an observed screen: reliability, cloud services, collaboration, content, integrations, certification, security, accountable support and trust accumulated over time.

The more resilient model may combine an open core with paid services, or accessible licences with genuinely portable formats. Before paying a subscription, customers and institutions will ask what data they can export, which APIs are documented, what happens when the contract ends and how much of the product's value truly depends on its vendor.

The digital Robin Hood does not fell the castle with one arrow: it forces the lord of the castle to explain why the bridge, maintenance and security still deserve a toll. That pressure is healthy as long as it does not confuse democratisation with impunity. The future of software will not be decided only by whoever recreates code fastest, but by whoever builds the most credible bargain between authors, communities and the people who depend on their tools. To assess that bargain in a real project, start with the working method and, when a decision needs context, describe the goal before choosing the technology.