In a June 2026 paper titled Cybersecurity is the True Frontier for Generative AI Success or Failure, researchers from CrowdStrike and several universities make a deliberately provocative claim. They argue that cybersecurity, and static analysis of software binaries in particular, is a more demanding test-bed for general AI progress than the fields that usually claim that title, natural language and computer vision. The reasoning is that a single malware sample can be viewed as a sequence of billions of tokens, that expert labeling is enormously expensive and slow, that skilled analysts frequently disagree on ground truth, and that all of this unfolds inside an adversarial environment where a well-funded actor is actively working to defeat whatever detection method you deploy. The full paper is available on arXiv (2606.28929) and is worth reading in its entirety.
We agree with the diagnosis. What we want to add is a conclusion the paper does not draw, because it is scoped to detection and analysis rather than to prevention architecture. If determining whether a file is malicious is genuinely this hard, and the paper makes a careful, well-cited case that it is, then the most durable defensive posture is one that does not require you to answer that question correctly before the file is allowed to reach a user. That is the premise behind Content Analysis, Disarm and Reconstruction, the CADR technology at the core of FileDNA. What follows walks through the four challenge areas the paper identifies and examines where CADR sidesteps the problem entirely, where it complements the AI-driven analysis the authors describe, and where the paper’s cautions apply to any prevention system as well.
The static analysis problem, stated plainly
Static analysis, as the paper defines it, is the task of inferring the properties of code without running it, an attempt to answer questions the halting problem forbids you from answering in absolute terms. For a Windows portable executable, that means reasoning about what a program will do purely from its bytes on disk. The authors are candid about the scale involved. They note that entire computer vision datasets are smaller than a single software binary, observing that ImageNet’s roughly 1.28 million images occupy only 17 gigabytes after resizing, while a single modern video game can exceed 40 gigabytes. Routine static analysis therefore means examining single samples an order of magnitude larger than whole datasets from other machine learning fields, and those samples do not parallelize the way conventional workloads do.
The expertise problem compounds the scale problem. The paper points out that a seasoned analyst can spend anywhere from hours to weeks dissecting a single binary, and contrasts this with radiology, where reading a scan takes from about a minute to sixteen minutes. Reverse engineering and disassembly are described as genuinely hard rather than merely tedious, because a linear-sweep disassembler will happily try to decode embedded data or jump tables as if they were instructions, and a single misdecoded byte can throw the entire disassembly out of sync with the real instruction stream. This is the reality that any detection-centric defense inherits.
Prevention-first design starts from a different question. Instead of asking how confidently we can classify this specific file before delivery, CADR asks how we can deliver a functionally equivalent file that is structurally incapable of carrying the malicious payload, whether or not we ever produce a verdict. Content Analysis inspects a file against a known-good specification for its type. Disarm strips the active and non-conforming elements that fall outside that specification. Reconstruction rebuilds a clean file from the legitimate content. The billions-of-tokens binary the paper describes as nearly impossible to analyze in real time is, under this model, a candidate for reconstruction rather than a subject for adjudication. That is the sense in which CADR sidesteps the hardest part of the problem the paper frames.
Tool use: complement, not replace
The paper’s third theme is the use of AI agents that orchestrate the many purpose-built tools of the security workflow, from command-line disassemblers to web APIs. The authors are measured here. They cite published work from Cisco and Google in which generative AI agents performed basic malware analysis inside fixed workflows, and they note both the promise and the limits, including one case where agents spent 46 minutes on a single, relatively simple sample that carried no significant countermeasures. They also point to autonomous tooling such as the DARPA Cyber Grand Challenge systems and the XBOW autonomous penetration tester as evidence that automation genuinely works, while being clear that these results are demonstrations rather than finished solutions. A recurring caution runs through this section: one cited study found that practitioners using generative AI performed no better than those without it, and that automation bias persists even after users are trained on the risks.
This is where CADR complements rather than competes with the direction the paper describes. Reconstruction removes the time pressure that the 46-minute figure represents. When a file is disarmed and rebuilt at the gateway, the deep analytical work the paper describes no longer sits on the critical path between the file and the user. The analyst’s agentic tooling can then run on the quarantined original at leisure, feeding threat intelligence and retrospective hunting rather than gating delivery. In practice, prevention and analysis become two loops running at different speeds: a fast reconstruction loop that keeps users safe now, and a slower analytical loop, increasingly AI-assisted along the lines the paper sketches, that develops understanding over hours or days. The paper’s honest accounting of what agents can and cannot yet do is, from this angle, an argument for not making the fast loop depend on them.
Long context and non-sequential structure
The fourth theme is the most technical, and the most interesting for anyone thinking about file structure. The paper explains that transformer self-attention scales quadratically with input length, which makes feeding an entire large binary to a model computationally punishing, while truncating the input throws away information that may be exactly where the malicious logic lives. Worse, the authors observe that malware authors deliberately pad binaries to make them larger and harder to analyze, turning file size itself into an obfuscation strategy. They also make a subtle point about structure: software contains jump instructions, so two instructions can be logically sequential even when their bytes are far apart, and two adjacent bytes may never execute in sequence. The spatial-locality assumption that makes transformers work well on text and images simply does not hold for compiled code.
CADR engages this problem from the structural side rather than the sequence-modeling side. The paper’s observation that malicious behavior hides in the gap between a file’s intended structure and its emergent, “shadow” structure is precisely the surface that Content Analysis targets. A specification-driven approach does not need to follow every jump or reconstruct control flow across a padded, non-contiguous binary. It needs to determine which parts of the file conform to the legitimate structure of its declared type and to discard the rest. Padding meant to defeat a sequence model is, to a reconstruction engine, simply non-conforming content that does not survive the rebuild. This is not a claim that structural analysis is easy, only that it is a different and more tractable question than whole-binary verdict generation. The paper’s own framing of shadow structure is, we would argue, the clearest articulation yet of why that difference matters.
Evaluation is hard, and it is hard for prevention too
The paper’s fifth theme is the one that deserves the most intellectual honesty in return, because it applies to any defensive approach, including ours. The authors argue that evaluation is the overarching difficulty in static analysis, driven by adversarially induced concept drift. Because adversaries refine their methods daily, any static benchmark begets overfitting to a snapshot of the world rather than to the real attacker-defender dynamic. They point out that concept drift creates strong temporal correlations that force time-based train-test splits, that recent data is far more valuable than old data for stopping current threats, and that intellectual property and data sovereignty constraints make it genuinely hard to assemble representative datasets. Software copyright, they note, prevents freely sharing a corpus of benign software, since that would amount to piracy, which starves researchers of clean ground truth.
We do not claim that CADR is exempt from this. A reconstruction engine still depends on accurate, current specifications for the file types it handles, and file formats evolve. The honest framing is that CADR changes the shape of the evaluation problem rather than eliminating it. Instead of asking whether a classifier’s decision boundary still holds against this week’s malware, which the paper shows can shift underfoot the moment an adversary adapts, the relevant question becomes whether reconstruction preserves legitimate functionality while excluding active content, across the file types in scope. That is a narrower and more stable target than chasing a moving malicious distribution, but it is not a free lunch, and any vendor claiming otherwise, ourselves included, should be read skeptically. The paper’s insistence on evaluation rigor is a discipline the prevention community should adopt, not an argument it gets to sidestep.
Deployment: latency, drift, and the analyst
The paper closes on deployment realities, and two of them speak directly to prevention design. The first is the demand for zero-shot and few-shot capability. The authors observe that a customer’s network changes day by day as users, systems, and software are added, and that new compiler versions and build tools shift the static analysis data distribution constantly, so any model will soon face an out-of-distribution environment even if its training data was perfect at the time. A prevention approach anchored to file-type specifications rather than to a learned distribution of known-bad samples has a structural advantage here: a novel malware family expressed through a conforming file format is handled the same way as a known one, because reconstruction does not need to have seen the specific threat before. This is the deployment analogue of sidestepping the detection problem.
The second is explainability. The paper argues that explainability is not optional in security, because analysts in a security operations center face hundreds or thousands of alerts a day and suffer real alert fatigue, and because understanding how an event occurred is necessary for remediation, not merely for confirmation. Here, prevention offers a genuinely different explainability story. When a file is reconstructed, the account of what happened is concrete and inspectable: these active elements were present, they did not conform to the specification, they were removed, and this is the clean file that was delivered. That is a narrative an analyst can verify without reverse engineering a binary, and it complements the AI-assisted explanation methods the paper surveys rather than replacing the analytical work that still needs to happen on the original.
Where this leaves us
The paper’s central claim is that cybersecurity is the true frontier for generative AI because it is so much harder than the fields where AI has recently triumphed, and its authors make that case with unusual care and a long list of primary sources. Nothing here disputes that. The additional point is architectural. If the verdict is this hard to reach, and the paper is persuasive that it is, then a defense that keeps users safe without waiting for the verdict is not a rejection of AI-driven analysis but a hedge against its remaining limits. CADR sidesteps the detection problem on the fast path, complements agentic and generative analysis on the slow path, and inherits its own share of the evaluation discipline the paper demands. The file is the front line. The most reliable way to hold that line is to make sure the file that reaches the user cannot carry the fight in the first place.
Reference: E. Raff, M. Ashkenazi, S. Samtani, D. J. Elkind, and S. Krasser, “Cybersecurity is the True Frontier for Generative AI Success or Failure“
FileDNA and its Content Analysis, Disarm and Reconstruction (CADR) technology are products of CyberQuay, Inc.