AI can suggest possible routes around patent claims, but a credible invent-around still requires experienced claim analysis, structured human invention, and clear conception provenance.
Seth Cronin | April 2026 | 8 min read
On April 7, 2026, Anthropic did something no AI company has ever done: they published a full system card for a model they will never release to the public.
Claude Mythos Preview, the engine behind Project Glasswing, is the most capable AI model ever built for finding software vulnerabilities. In testing, it discovered a 27-year-old bug in OpenBSD – an operating system specifically engineered for security. It found a 16-year-old flaw in FFmpeg that automated testing had scanned five million times without catching. It produced working exploits for Firefox at a rate of 181 out of several hundred attempts, where the previous best model managed two.
And Anthropic looked at what they’d built and said: this can’t go public. Not now. Maybe not ever.
About 50 organizations got access. Everyone else got a blog post.
I’ll be honest: that wasn’t on my 2026 bingo card. Every other frontier model has gone to general availability. This one might never. The most powerful model ever created, and a very small in-crowd will have access to it, able to totally dominate infrastructure, markets, and probably knowledge work writ large – all in the name of safety.
I don’t entirely disagree with the safety argument. AGI-class capabilities in the wrong hands would be catastrophic, even before you get into the x-risk discourse. But what’s happening in cybersecurity right now is a preview of what’s coming for intellectual property. And nobody in the IP world is talking about it.
Here’s the core idea behind Project Glasswing: software can contain vulnerabilities that human reviewers miss, and AI can sometimes surface them quickly. The OpenBSD SACK bug sat in security-critical kernel code for 27 years. Professional security teams. Five-figure audit budgets. Decades of eyeballs on the code. None of them caught it. Mythos reportedly found it overnight while an engineer with no formal security training slept.
Glasswing’s premise is uncomfortable but useful: defenders should test their own systems with the same class of tools an attacker might use. AI adds speed and breadth to that testing. It does not eliminate the need for experts who can distinguish a real vulnerability from a confident false positive.
Now replace “software vulnerability” with “patent claim limitation.”
Many patents contain claim limitations that create design-around opportunities. Once identified, a limitation may suggest another way to pursue the same commercial objective. But an apparent route around the words is not automatically a noninfringing implementation. Claim construction, dependent claims, related patents, prosecution history, and the doctrine of equivalents can all change the answer. Patent professionals know this intuitively. It’s why design-around work is an invention and claim-analysis discipline, not a keyword exercise.
But here’s what changes when AI enters the picture: scale and speed.
An AI system can propose dozens of alternative implementations in the time a human team once needed to develop one or two. That changes the speed of the first pass, but not the reliability of the conclusion. A language model is generating hypotheses from claim language. It is not construing the claim, applying prosecution history or the doctrine of equivalents, checking the rest of the patent family, or deciding whether a real implementation is technically and commercially viable.
That distinction matters. A fast answer can reveal a promising direction. It is not a freedom-to-operate opinion, and it is not a completed invent-around strategy. The difference becomes clear in a real example.
Let’s take a real patent. Not a hypothetical. Not a textbook example. A patent granted in April 2026.
US 12,597,427 B2 – “Configurable Natural Language Output”
Granted: April 7, 2026
Assignee: Amazon Technologies, Inc.
Filed: June 1, 2023
This patent covers a system that changes how an AI assistant talks based on who owns the device it’s running on. A hotel chain deploys an Alexa in every room, and the assistant speaks with a hospitality-focused warmth. A toy company uses the same underlying platform, but the voice is playful and kid-friendly. The “personality” is tied to the organization and the device, selected by a trained machine learning model.
Here’s the heart of it – Claim 1, the broadest independent claim, stripped to its structural bones:
Element A: Receive a natural language user input
Element B: Determine device data indicating an organization associated with the device
Element C: Use a trained model to select from a plurality of language generation profiles
Element D: The selected profile is associated with the organization
Element E: The model was trained on examples that include: input + device data + language generation profile
Element F: Generate output using the selected profile, where the output includes words corresponding to the profile’s style
Literal infringement generally requires every limitation of a claim, but whether a product meets a limitation is a claim-construction question, and the doctrine of equivalents may still matter. The following are candidate design-around directions, not clearance conclusions.
The claim requires a trained model that selects from profiles based on training examples containing input, device data, and profiles. What if you don’t use a trained model at all?
Candidate path: A rule-based configuration system. The organization sets up a simple mapping: “If device_owner = Marriott, use style = hospitality.” No machine learning. No training examples. A lookup table. On a first reading, that architecture may avoid the trained-model requirements in Elements C and E. A reliable conclusion would still require review of the complete claim set and the actual implementation.
The claim requires selecting from among a plurality of language generation profiles – discrete, pre-existing style templates. What if profiles don’t exist as discrete entities?
Candidate path: Dynamic style generation at inference time. Instead of selecting from a menu of profiles, inject style instructions directly into an LLM’s system prompt at the moment of each request. “You are a concierge at a luxury hotel. Speak warmly and formally.” The style is composed on demand rather than retrieved from a pre-built collection. That may provide a route around the profile-selection language, subject to how a court would construe the claim and how the system is actually built.
The claim requires device data indicating an organization associated with the device. What if the style trigger isn’t an organization?
Candidate path: User-preference-based personalization. The system detects the user’s speaking style, sentiment, or explicit preference settings (“I prefer formal responses”) and adapts accordingly. The style follows the user, not the device owner. That difference may avoid the organization-device relationship recited in Elements B and D, but it must be tested against the full patent position.
Combine all three. Build a system where:
At a high level, the combination appears to preserve the commercial objective while changing several requirements recited in the claim. That is the beginning of an invent-around analysis, not the end. The complete claim set, related family members, prosecution history, doctrine of equivalents, prior art, and proposed technical implementation still need to be evaluated.
Total time to generate four candidate directions: about fifteen minutes.
Claude produced those candidates. It did not complete an invent-around.
It performed useful first-pass work: decomposing the claim, highlighting pressure points, and proposing alternative architectures. The professional work begins where that output stops. Someone still has to interpret the claims, test technical feasibility, review dependent claims and related patents, consider prosecution history and equivalents, search the broader landscape, and determine whether the proposed implementation actually creates operating room.
The dangerous failure mode is artificial confidence. AI can state that a limitation has been avoided with the same confidence it uses when it has overlooked a dependent claim, misread a term, or proposed an architecture that is technically impossible. More model capability creates more candidate routes. It does not turn those routes into legal or technical conclusions.
At ipCapital Group, ipNavigation uses a proven 18-point framework to guide technical experts through a disciplined invent-around process. We analyze the claims, identify the actual design constraints, and use eighteen distinct invention strategies to generate alternatives from human experience rather than relying on one model’s most statistically convenient answer.
ipCG inventors and the client’s technical experts then evaluate, combine, and extend those alternatives. The result can be a credible path around blocking IP, a stronger defensive position around the client’s own patents, and new potentially patentable subject matter for counsel to assess. Just as importantly, the process preserves who conceived what as the ideas develop.
Turn candidate routes into an invent-around strategy
See how ipCG’s 18-point ipNavigation process helps teams navigate blocking patents, strengthen their own claims, and develop human-conceived alternatives.
Fair question. A reader asked whether this is really different from what patent attorneys have always done – just finding prior art faster. Here’s why I think it’s different this time, and I don’t say that lightly.
The bottleneck argument cuts both ways. A smart critic pointed out that patent offices and courts run on human timescales. An AI that generates a thousand novel claims per hour doesn’t help when the system processes twelve per month. True. But that’s the defender’s bottleneck, not the attacker’s.
In cybersecurity, CrowdStrike’s CTO put it bluntly: “The window between vulnerability discovery and exploitation has collapsed – what took months now happens in minutes with AI.” A related asymmetry applies to IP. A competitor can use AI to generate candidate routes around a portfolio at machine speed. Patent owners still have to evaluate those candidates, develop technically credible responses, preserve human conception, and file on calendar time. The speed advantage is real, but it is an advantage in generating possibilities, not proof that any particular route works.
Generating a candidate is constraint work; completing an invent-around is invention strategy. A model can recombine known alternatives against a list of claim limitations. A defensible path may require a genuinely different architecture, material, process, or control scheme that a technical team must conceive and develop. The most valuable outcomes are not merely ways to omit a word from a claim. They are new technical positions that solve the business problem and may themselves support patent filings.
The combinatorial advantage is real. AI can evaluate substitutions, omissions, and alternative combinations across a large portfolio faster than a human team. That makes it a useful search instrument. The value comes from pairing that breadth with experienced claim interpretation, structured invention methods, and technical judgment so the team can distinguish a plausible sentence from a viable strategy.
The patent-rights issue cuts in the opposite direction from the usual AI hype. The USPTO’s November 2025 revised guidance for AI-assisted inventions says AI-assisted inventions are judged under the same inventorship standard as any other invention. AI systems may generate ideas, but they are tools, and only natural persons can be inventors. The touchstone remains human conception: a specific, settled idea of the complete and operative invention.
Using Claude does not automatically make an invention unpatentable. But if Claude supplies the complete claimed solution and the humans merely recognize it or reduce it to practice, there may be no natural person who can honestly establish conception of that claimed subject matter. An AI-only invention cannot be cured by simply putting a person’s name on the application. That can leave the claims with a fundamental inventorship problem.
The answer is not to ban AI. It is to preserve human provenance. Document who framed the technical constraints, who proposed each alternative, why the team selected or rejected it, how the concept changed, and which human conceived the features that later appear in the claims. A facilitated invention process does more than produce ideas; it creates a defensible record of where those ideas came from.
The cybersecurity industry went through its own reckoning. The lesson, painfully learned: you can’t just build walls. You have to continuously test them.
Before Glasswing, the security model was largely reactive. Wait for a breach. Patch the hole. Pray you got there in time. Glasswing inverts this: actively probe your own defenses with machine-speed assistance, then have experienced people determine what is real and what to do about it.
The IP world is still in the “wait for someone to infringe” era. Most companies treat their patent portfolios as static assets – filed, granted, maintained, enforced when necessary. Almost nobody proactively stress-tests their own claims the way a red team stress-tests network security.
That is where AI belongs: inside a disciplined claim-testing and invention process, not in place of one. ipNavigation combines AI-assisted analysis with an 18-point human-facilitated framework to identify:
AI finds candidate pressure points quickly. The facilitated team determines which ones matter, invents credible alternatives, and builds the record needed to act on them.
If you’re a CTO, general counsel, or IP strategist reading this, here’s my honest recommendation:
1. Use AI for hypothesis generation, not clearance. Let it decompose claims and propose alternatives quickly. Treat every answer as a lead to investigate, not a conclusion that you are free to operate.
2. Put every candidate through claim and portfolio analysis. Review the full claim set, related family members, prosecution history, equivalents, prior art, and the proposed product architecture. An apparent gap in one independent claim may disappear when the surrounding position is considered.
3. Keep human inventors involved from the beginning. Use structured invention sessions to challenge, combine, and extend the machine-generated candidates. Record the human contributions while they happen so potentially patentable subject matter has a clear conception history.
4. Use the results in both directions. For blocking competitor IP, develop several technically credible routes around it. For your own portfolio, anticipate those routes and build continuation claims or new filings that make the position harder to navigate.
AI makes candidate routes cheap and abundant. It does not make interpretation, technical experience, facilitation, or invention provenance less valuable. It makes those scarce capabilities more important.
A Claude response can show a possible path around a patent. ipNavigation tests whether that path survives claim analysis, turns it into a workable technical alternative, and helps human inventors develop related subject matter that may be protectable. That is not a service AI eliminates. It is the layer that keeps AI output from becoming an expensive false positive.
For patent owners, the same process exposes gaps before competitors exploit them. For teams blocked by another company’s IP, it creates multiple human-vetted routes instead of one confident machine answer. Machine speed is useful for generating options. Human judgment and conception are what turn those options into strategy.
If you use AI to invent around patents, tread lightly. It may help you find the start of a route. Without experienced claim analysis and a human invention process, it can also give you false confidence about infringement and weaken the provenance of patent rights you hoped to create.
Need to understand what this means for your portfolio? ipCapital Group helps leadership teams turn patent landscapes into practical decisions about filing strategy, competitive diligence, M&A readiness, and monetization.
Start a conversation with ipCapital Group or download the IP Strategy Playbook.
Work with ipCapital Group
From invention to monetization, our team has guided 2,000+ engagements across the full IP lifecycle. Start with a free 30-minute discovery call.
Written by
Seth Cronin