When AI Clones Software, What Counts as Original?

Developers using AI tools to clone popular paid software, most notably Adobe, have opened a debate that reaches far beyond software pricing. Some people call the practice theft, while others welcome free open-source tools that give more users access to capable software.

Matt Wood, Senior Director of Cyber Security Architecture, Operations, and AI Security at ABBYY, has investigated the concern. His view connects software cloning to a larger question about AI development: when a system can reproduce the shape of an existing product, where does originality begin, and where does risk enter the picture?

Music faced a similar argument decades ago. In 1996, Fatboy Slim built “Rockafeller Skank” from a guitar riff lifted from a 1960s record that few people had heard in decades. The Chemical Brothers looped breakbeats from forgotten funk records into stadium-filling anthems, while The Prodigy turned sirens, film dialogue, and obscure vinyl into something that sounded like nothing before it.

Sampling did not only lower the cost of making music. It changed who could make it and forced the industry to ask what actually counts as original. AI is asking enterprises the same question, but the stakes now include operational outcomes, compliance exposure, and the trust customers place in every business decision.

AI Can Reproduce Fluency Without Providing Judgment

Large language models make it easy to remix existing knowledge. They can draft a contract clause, summarize a report, or produce a plausible answer from the text they can access. That capability has real value, but replication is not discovery, and confusing the two is where many enterprise AI projects stall.

Generative systems can produce work that looks finished. They imitate tone, structure, and confidence with extraordinary fluency, yet they cannot guarantee on their own that the result rests on the right evidence, connects to the right process, or reflects sound judgment.

A 2024 Science Advances study found that AI assistance improved the perceived quality and creativity of individual stories, especially for less creative writers. At the same time, the full set of stories became more similar. The output improved at the individual level but converged at the collective level.

That is the enterprise risk in miniature. More production does not always mean more originality. It can mean more of the same ideas, expressed more smoothly, while the real advantage comes from knowing what matters, what to trust, and what deserves action.

Most businesses do not lack language. They lack accessible, reliable context. What a company knows is scattered across contracts, invoices, onboarding forms, claims, emails, reports, scanned PDFs, and the workflows that move those documents from one decision to another.

Enterprise AI Needs Context Before It Needs More Output

The pressure to find better context is arriving alongside waste. A survey of 2,018 IT heads at big firms found that enterprises waste 28 percent of their software budgets on unwanted upgrades, pointing to a lack of leverage in the way businesses buy and use software.

ABBYY’s answer is PHOENIX Plus, a native generative AI layer built into the ABBYY Vantage platform. It is part of PHOENIX, ABBYY’s portfolio of proprietary and curated AI technologies, and works with PHOENIX Core, the deterministic foundation built around optical character recognition, layout analysis, classification, extraction, and document-specific intelligence.

PHOENIX Plus adds generative models selected, optimized, and hosted by ABBYY. The system converts documents of any complexity into structured, context-rich data that AI-powered automation can consume, while supporting decision recommendations, analysis, and reasoning. ABBYY lists 99% accuracy on the first pass for PHOENIX Plus.

Vantage users can apply zero-shot extraction to unseen, long-tail, and highly variable documents by using prompts to define the output schema instead of labeled training data. Pre-engineered prompts reduce the work of designing and testing prompts from scratch, while validated data can be enriched through free-form text interpretation and inconsistency checks.

“Enterprises today want the benefits of generative AI in their document workflows but require the compliance and auditability that enterprise-grade technology provides,” said Bruce Orcutt, Chief Product and Marketing Officer at ABBYY. “PHOENIX Plus was built specifically to close that gap.”

PHOENIX Plus is hosted by ABBYY, keeping data in the same region as a customer’s Vantage data. Customers add it to their Vantage subscription at a fixed annual cost, without a separate model vendor contract, infrastructure to provision, or token-based billing to manage. Vantage also supports connections to OpenAI, Google Gemini, Anthropic Claude, Mistral, and custom LLM connections through its bring-your-own-model option.

Software Factories Change the Security Problem

AI cloning raises questions about originality, but AI-generated software creates another urgent problem: security. “We don’t have a cybersecurity problem. We have a software quality problem,” Jen Easterly, former director of the Cybersecurity and Infrastructure Security Agency, said at the BlackHat security conference in 2024.

That statement has gained force as development moves toward software factories, where specifications go in and agents write, revise, test, and ship code. Humans set intent, review outcomes, and govern the system, but the primary author of much of the code is no longer a person who can be trained, reviewed, or held accountable in the usual way.

Traditional application security treated security as a gate or audit after development. “Shift left” moved finding and fixing issues closer to the moment of creation, which helped people write safer code and avoid dependencies with known vulnerabilities. That approach worked when humans wrote code at linear rates and attackers lacked the time and tools to chain lower-severity findings.

Those conditions have broken. Analysis of CVE.org and National Vulnerability Database data shows that new security issues increased dramatically in 2026, while autonomous attack agents are no longer a research curiosity. The attacker is no longer resource-starved, and deferring the backlog has become an active liability.

Even the best models produce fully correct and secure backends well under half the time. BaxBench showed the strongest models reaching roughly 60% functional correctness, with a large share of those correct solutions still insecure. Development velocity has risen by an order of magnitude, but the quality floor has not.

Why “Shift In” Moves Security Into Creation

“Shift in” means putting security context, boundaries, and feedback inside the agentic creation loop before code is emitted. Guardrails sit on the agents, security context lives inside them, and organizations burn down the software already shipped.

Agentic AppSec is not simply a better scanner or “shift left” with new branding. Detection must combine deterministic analysis, frontier intelligence, and broad context about how vulnerabilities behave in real systems. Remediation must be intelligent, and prevention must operate inside the software factory.

An independent judge must also test the running software or pre-production candidate like an attacker, across the agentic, web, and API layers. Authorization flaws, business logic errors, and agent-specific behaviors may only appear when someone probes a working system as an adversary would.

That independence matters because the generator cannot validate itself. A model grading its own homework is not a security program; it is a conflict of interest with a dashboard.

Agents can select packages that contain malware or invent package names that do not exist. Attackers can register those names first, allowing the next agent to install the payload through a technique known as slopsquatting. Research into the agent skills ecosystem found that roughly a third of skills carried security issues, including prompt injection, credential theft, and confirmed malicious payloads. Toxic flows and poisoned MCP servers have also been demonstrated by attackers.

Humans still matter, but their role must move from inside the loop to on the loop. They set the bar, govern agents, audit outcomes, decide what good looks like, and hold the system to it. That shift creates the AI Security Engineer, even as most organizations remain outside true software factories and continue writing specifications, prompting agents, and iterating.

Stay connected via Google News

Follow us for the latest AI News