Can You Actually Keep It Secret?

Can You Actually Keep It Secret?

Trade Secrets When Reverse Engineering Gets Cheap

In a recent piece I argued that agentic coding is draining your moat. When a capable team with the right tools can reach feature parity in days, the advantage of shipping first stops being defensible and becomes table stakes.

That argument has a companion. Many companies, having watched the speed moat erode, reach instinctively for another one: lock the crown-jewel technology away as a trade secret. The instinct is sound. It also deserves more scrutiny than it usually gets, because the same force that drained the first moat, cheap and capable AI, is quietly eroding this one as well. It is making reverse engineering far easier than it used to be.

My earlier article on the agentic-coding moat is available at slwip.com: Agentic Coding Is Draining Your Moat.

Two questions that get blurred into one

When a company decides to protect something as a trade secret, it is really answering two very different questions. Most teams blur them together. Keeping them apart is the whole game.

The first question is whether you can keep the information secret at all. Do you have the measures and procedures in place to qualify for, and to hold, trade secret protection? The second question is whether the technology can be reverse engineered. The first is largely legal and procedural. The second is technical. You need a good answer to both, because a strong answer to one cannot rescue a weak answer to the other.

Question one: can you actually keep it secret?

Trade secret protection rests on two requirements. The information must derive independent economic value from not being generally known, and the owner must take reasonable measures to keep it secret. Both the Defend Trade Secrets Act and the Uniform Trade Secrets Act are built on that pairing. The Fourth Circuit revisited the reasonable-measures standard in November 2025 in Samuel Sherbrooke Corporate Ltd. v. Mayer, holding that, at the pleading stage, a confidentiality agreement alone can be enough to show reasonable measures. The bar to get into court is not high. What it takes to actually prevail is another matter.

This is where your measures and procedures earn their keep. NDAs, access controls, segregation of sensitive materials, exit interviews, and clear confidentiality policies all do real work. That work is aimed primarily at improper acquisition: theft, breach of a duty of confidence, and the growing problem of employee mobility.

Employee mobility is the risk I see rising fastest. People move between competitors, and they carry know-how with them. Recent litigation reflects it, and the clearest examples are now in software. In Motorola Solutions v. Hytera Communications, a competitor recruited several Motorola engineers who, on their way out the door, carried off thousands of proprietary documents and the source code behind Motorola’s digital radio technology, which Hytera then used to build a competing product line. The Seventh Circuit affirmed the resulting judgment in 2024, in an opinion that noted Hytera had effectively conceded the blatant theft of trade secrets and copying of proprietary code, and the damages ran into the hundreds of millions. Hytera later pleaded guilty to a criminal trade-secret conspiracy charge under the Economic Espionage Act. Cases like that turn heavily on whether the original owner had genuine procedures in place.

Good measures are what let you act when someone walks out the door with your secret. The same procedures can also be undone from the inside, and the AI era has opened a new way to do it. In Trinidad v. OpenAI, an early 2026 decision from the Northern District of California, a court dismissed a DTSA claim where the plaintiff had developed her asserted frameworks by feeding them into ChatGPT. Putting the material into a public tool whose terms granted the provider rights over the inputs meant she had disclosed it to a party under no obligation to keep it secret, which extinguished any trade secret. It is a thin decision, a district court dismissing a pro se plaintiff, so do not read too much into it. Read it instead as an early marker: every employee who pastes proprietary code or design notes into a consumer AI tool may be doing quietly what a departing engineer does on the way out the door. If employee mobility is your principal exposure, trade secret protection backed by disciplined procedures can be exactly the right tool, and that part of the analysis is very much alive.

Question two: is it reverse engineerable?

Here is the point clients hear least often. Trade secret law has never protected you against reverse engineering. Reverse engineering is a recognized, lawful means of acquiring a secret. If a competitor can buy your product on the open market and work backward to your innovation, the law does not stop them, and the moment they succeed your secret is gone. A patent protects you even against a competitor who independently invents or reverse engineers the same thing. A trade secret does not. That single distinction is doing more work today than it ever has.

What has changed is not the law but the cost. Reverse engineering used to be slow, expensive, and uncertain, often expensive enough that secrecy held in practice even when it was vulnerable in theory. That economic friction was the real moat. AI removes much of the friction, for software and increasingly for physical devices. The uncomfortable part is that modern models are remarkably good at reproducing a result given enough examples of inputs and outputs. They do not need your source code. If the behavior of your system is observable, a model can often learn a functional equivalent of it, close enough to erase your advantage. A black box stops being black once someone can interrogate it at scale.

Whether a given technology can be reverse engineered is a technical question, not a legal one. It has to be answered by the people who understand the system, and the analysis has to look honestly at every surface the technology exposes. The user interface is only the start. Every surface a third party can observe or call, from the APIs and SDKs down to the wire protocols and file formats, is a place where secrets can leak or be probed.

Two concrete software examples sharpen the line. Some innovations really are hard to reverse engineer. They tend to live server side, operate over high-dimensional inputs, and never expose their internals. Consider the feature engineering and signal weighting behind a fraud or risk score that blends hundreds of inputs into a single decision. An outsider sees a number, not the recipe, and cannot gather enough clean input-output pairs to back it out. The training recipe for a proprietary model is similar: the data mixture, the cleaning and curation pipeline, the hyperparameter choices. None of that leaves a usable fingerprint in the outputs. This is the modern equivalent of the formula locked in the vault, where the value lives in a process rather than in the visible result.

Other things are exposed the moment they ship. Anything delivered to the client can be inspected, whether it is browser JavaScript, a mobile binary, or firmware. More important still, any function whose inputs and outputs are both observable and simple enough that a model can learn the mapping is fragile no matter how elegant the implementation. A client-side transformation, a deterministic image or audio effect, a pricing or matching rule a competitor can probe with queries, all of these are what AI-assisted reverse engineering handles with ease. When the value lives in observable behavior, secrecy is the wrong tool. Patent it.

A technical screen: three feasibility tests

When a client asks whether to patent something or keep it secret, I find it useful to run the technology through three questions on the reverse-engineering side before deciding:

1. Output discernibility. From the product’s observable behavior alone, how much of the innovation can a competent observer, or a model, infer? The more the value shows up in the output, the weaker secrecy becomes.
2. Interface exposure. What does the technology expose at its boundaries? Count the APIs, the CLIs, the MCP servers, the SDKs, the wire and network protocols, the file formats, and the webhooks. Each one is a surface a third party can study.
3. Replication cost. Even where the innovation is theoretically discernible, how expensive is it to actually reproduce, and does that cost still hold now that AI has lowered it?

This is a joint decision, not a legal one

The mistake I most want in-house teams to avoid is treating this as a question legal can answer alone. It cannot. Question one lives with legal. Question two lives with the technologists. The decision needs both groups in the same room. The most useful exercise I know is to put your own engineers in the hypothetical shoes of a well-resourced competitor. Hand them the shipped product and ask, honestly, whether they could reverse engineer the thing you are trying to protect, and how long it would really take now that they have these tools. Their answer, not the lawyer’s instinct, should drive that half of the decision.

None of this is a case against trade secrets. For the right technology, genuinely opaque, genuinely costly to reproduce, and guarded by real procedures, a trade secret remains one of the most powerful protections available. It can last indefinitely, and it requires no filing at all. The point is narrower. A trade secret is only ever worth as much as the weaker of its two answers. Perfect procedures cannot save a secret that is trivially reverse engineered, and an un-reverse-engineerable secret still needs the procedures to survive a departing employee.

What to do about it

AI-era development often forces this call sooner than you would like, with less certainty than you would like. That argues for keeping your options open. Filing a provisional application early buys a twelve-month window to run both analyses properly, the legal one with counsel and the technical one with your engineers, without forfeiting patent rights while you decide. If the assessment comes back that the technology can be genuinely kept secret, you can let the provisional lapse. If it comes back the way it increasingly does, you have preserved the protection that actually survives reverse engineering.

Trade secrets are not dead. The set of technologies that can realistically be kept secret is narrower than most companies assume, and it is narrowing faster than they have noticed. The discipline now is straightforward to state and harder to practice. Ask both questions early. Answer each with the people who actually know. Stop assuming a secret will keep simply because you have decided to call it one.

This article is provided for general informational purposes and does not constitute legal advice.