The Secret Is Out: AI Reverse Engineering and the Patent vs. Trade Secret Decision

For decades, the choice between using a patent to protect your software innovation and keeping it as a trade secret rested on an assumption: reading someone else's compiled code is slow, expensive and rarely done. That meant the odds of a competitor studying your code, let alone copying it, were low. AI agents have now broken that assumption. 

In April 2026, the UK AI Security Institute reported that a frontier model solved a reverse-engineering challenge in 10 minutes and 22 seconds, at a cost of US$1.73. A human expert with professional tools needed about 12 hours and, one assumes, billed for all of them. Six months later, an open-source project called REA ("Reverse Engineer Anything") went from under 10,000 GitHub stars to over 100,000 in a single week by connecting ordinary AI coding agents to professional decompilers. Its own sample prompt invites a developer to point an AI agent at an app, work out how a feature functions without the source code, and build a version of it. 

REA is one early, visible example of a fast and low-cost reverse engineering capability that is now available to any developer, and it will not be the last. When reverse engineering costs almost nothing, the patents-versus-trade-secrets decision changes. 

In short: for any innovation a competitor can easily copy, secrecy should now be treated as temporary from the day of launch. At the same time, the tools that erode secrecy make software patents easier to enforce, and much more valuable where they are available. 

The old calculus and its hidden premise 

The textbook trade-off between patents and trade secrets is familiar. A patent gives a strong, time-limited monopoly in exchange for public disclosure, and it costs time and money to obtain and enforce. A trade secret costs little up front and can last indefinitely, but it protects only against misappropriation. It is important to note: it offers no protection against a competitor who independently develops or reverse engineers the same thing.  

For a lot of software innovations, companies chose secrecy anyway. Source code is not shipped; compiled code is hard to read; and a competitor willing to spend weeks of specialist time decompiling a feature usually had better things to do, like building their own software. What protected shipped software was practical secrecy, priced in engineering hours. For decades, the most effective trade secret protection in software was tedium. 

What changed? 

Decompiler tools such as Ghidra, IDA and Hopper have existed for years, but using them well required specialized knowledge and expertise. An AI agent can now drive them, read the output, trace a feature from the user interface down to individual functions, and explain what it found in plain language. Tools such as REA package this as a one-command install for mainstream coding agents, covering native binaries, Android apps, Electron and JavaScript apps, .NET assemblies and websites. 

This has changed who can reverse engineer, and how often, opening up the field of play. A task that once required a team of specialists and a budget can now be run by any developer, on any competitor's shipped product, as casually as a web search. Software vendors have noticed: some are already rewriting their licence terms to prohibit AI-assisted reverse engineering. Those clauses may help at the margins, but a contract binds only the parties to it, and its enforceability varies by jurisdiction. 

The practical moat around reverse engineering shipped code has drained. The legal one – trade secret protection – may follow. 

Is it still a secret at all? 

The conventional view is that reverse engineering defeats a trade secret only when someone actually does it. Until then, the owner keeps its rights against anyone who takes the secret by improper means, such as a departing employee or a breach of an NDA. 

But in several jurisdictions, the definition of a trade secret itself depends on how hard the information is to obtain: 

Now, courts are unlikely to accept that argument wholesale any time soon. "Readily" has never meant "possible", and a well-run secrecy program still counts for a great deal. But the cost of extraction will only fall, and a protection strategy that depends on a judge treating a two-dollar extraction as if it still took two months should be revisited now. 

Patents just got easier to police 

One long-standing reason companies undervalued software patents was detectability. A patent is only as good as your ability to prove infringement, and the evidence was locked inside a competitor's black box. Many software patent owners suspected infringement but could not justify the cost of proving it before litigation. 

The AI tools that dissolve practical secrecy also ease that problem. An AI agent that can trace how a competitor's app implements a feature can, at least in principle, produce a claim chart's worth of technical evidence at a fraction of the former cost. (Whether and how such analysis may be done in a given case raises its own licence and copyright questions, which I take up in a companion article.) 

The same shift pushes in two directions: 

The disclosure cost of patenting also shrinks. If a competitor can read your shipped implementation with an AI agent, publishing a patent specification reveals little that the product has not already given away. 

What if your software innovation is not patentable? 

Under current laws in many jurisdictions, a patent may not be available for every software innovation, which leaves many innovators exposed. In the United States, the USPTO's precedential Ex parte Desjardins decision, and the examiner guidance that followed, have made patent examination friendlier to specific technical improvements, but neither binds the courts. In Canada, the substance has barely moved despite new dressing. The Federal Court of Appeal in Benjamin Moore (2023) declined to adopt a dedicated test, and CIPO's March 2026 guidelines continue to hold that an algorithm run on a computer in a well-known way remains unpatentable. 

That leaves a growing class of innovations in a squeeze. A competitor can copy, buy, download or call them, so secrecy will not hold; and if the innovation in their software does not improve the computer itself, a patent may be hard to obtain. For those cases, neither traditional IP tool works well, and protection has to come from elsewhere: speed to market, keeping the valuable part of the innovation on a server (out of reach of tools like REA, for now), proprietary data and network effects, and contract terms with customers and partners. 

A new sorting rule 

The old question was, "is this patentable, and can we keep it secret?" The more useful question now is "what can a competitor observe once they buy, download or call it?" 

Two cautions. First, the server-side category is smaller than it looks: these AI tools also capture network traffic and can rebuild an application's API from a recorded session, so a surprising amount of back-end design leaks through the interface. Second, AI models deployed on-device move from the bottom row to the top one. Model files shipped with an app can be inspected like any other artifact. 

What to do now 

For in-house counsel and CTOs: 

For innovators: Where your innovation is patentable, the cost-benefit of filing has moved in your favour. Where it is not, plan for it to be copied and build your advantage elsewhere. If investors ask what stops a large incumbent from copying your app, "they would have to reverse engineer it" is no longer an answer. 

For patent agents and attorneys: When drafting software claims, aim for infringement you can prove from the outside. Claim the client-side and interface-level steps a reverse-engineering agent could observe, and as always avoid claims that can be infringed only by steps hidden on a server you will never see. Detectable claims were always good practice; they are now slightly easier to write and more valuable than ever. 

The bottom line 

Trade secrets are not dead. Server-side systems, data and training pipelines remain natural candidates, and secrecy still beats a patent that would not survive a validity challenge. But for innovations a competitor can buy, download or call, the economic premise behind choosing secrecy has collapsed, and the legal premise may follow. Companies that revisit the decision now, while grace periods and filing windows are still open, will be in a better position than those that discover the change in a courtroom. 

This article provides general information only and is not legal advice. For advice on your specific situation, please contact Paul Horbal or a member of the Marks & Clerk Patents team.