Enterprises have spent the past two years celebrating how much faster AI agents can write software. Nir Valtman, chief executive of the application security firm Arnica, is more interested in a quieter question: who is accountable for all of it, and can you prove it?
Speaking at OWASP Global AppSec EU in Vienna, where Arnica has been a Diamond sponsor since the company’s earliest days, Valtman described what he calls the “governance gap,” the widening distance between the speed at which AI coding agents produce code and the ability of security and engineering teams to review, own, and account for it. His central claim is that this gap is not an abstraction. It is already showing up on balance sheets, in staffing decisions, and, increasingly, in the questions auditors ask.
Valtman locates the cost in three measurable areas. The first is remediation backlogs. When agents generate code faster than security teams can triage it, unresolved findings pile up, and “each item sitting in that backlog is a quantifiable risk-carrying cost, not just a to-do item,” he said. A backlog, in his framing, is unpriced risk accumulating on the books.
The second is reviewer burnout and attrition. When every pull request is treated with identical scrutiny regardless of the risk it actually carries, an organization’s most senior engineers end up spending their hours on low-value triage rather than architecture or mentoring. That misallocation, Valtman argues, is a direct and often underestimated productivity cost, and it compounds when those engineers eventually leave.
The third is the one he says auditors are now raising explicitly: accountability exposure. “If a security incident traces back to AI-generated code and your organization can’t produce evidence that a human or a verified control reviewed it,” he said, “that’s not just an engineering problem, it’s a governance and potentially a regulatory problem.” His recommended instrumentation for leaders is concrete: track time-to-triage, monitor backlog growth rate, and, above all, be able to produce auditable evidence of review for any given piece of production code.
Underlying the gap, in Valtman’s telling, is an economic model that no longer fits. Traditional application security is reactive: code is generated, then scanned, then remediated, then scanned again. “You pay to generate code, you pay again to scan it after the fact, you pay a third time to remediate whatever the scan finds, and then you pay again to scan the code,” he said. Each pass consumes tokens, engineering time, and release velocity.
Arnica’s counter-proposal is what the industry has begun calling secure-by-default: collapsing those repeated passes into one. Rather than scanning everything repeatedly, Valtman says organizations should “instruct the agents how to build it securely as part of the product specifications, so it will build it the right way the first time.” The saving is in compute cost, review time, and delayed releases avoided.
He is candid that this is hard to scale. Secure-by-default asks developers to make changes they are “not measured on,” which is a governance problem as much as a technical one. Arnica’s wager is that a developer-native workflow, one that meets engineers inside the tools they already use, is what converts a good idea into actual adoption.
For a European executive audience, the sharpest part of Valtman’s argument is regulatory. He contends that both the EU AI Act and NIS2 are pushing organizations in the same direction: away from self-attestation and toward demonstrable accountability. The AI Act’s risk-based approach, he notes, increasingly expects organizations to show, not merely claim, that meaningful human oversight exists for AI systems operating in sensitive contexts, and AI agents shipping code that touches critical infrastructure or personal data plainly qualifies. NIS2, meanwhile, raises the bar on incident accountability and supply-chain risk management for essential and important entities.
The common thread, he argues, is a single question regulators and auditors will keep returning to: “can you prove a human was meaningfully engaged at the right point in the process, and can you prove it after the fact, not just assert it?” His answer is the review record, an artifact capturing what a reviewer actually saw and what they did about it. That, he says, is what turns “we believe we have oversight” into “we can demonstrate we had oversight.” His advice to compliance and security leaders is to start treating that evidence as a first-class deliverable rather than a byproduct.
Notably, Valtman is clear that the buyer for Arnica’s code review product is not, in the first instance, the security chief. The pain it addresses is reviewer fatigue, review cycle time, and where attention gets routed, which makes a VP of Engineering or CTO the natural purchaser. “The risk reduction is downstream of solving the engineering problem,” he said. Practically, the system classifies files before any human opens them, sending roughly 70% to a “Skim” tier so that reviewer focus lands on the 30% that warrants it. The security benefit, in his view, follows from fixing the engineering bottleneck rather than the other way around.
That positioning has drawn analyst attention. Arnica points to inclusion in Forrester’s application development and security landscape for the second quarter of 2026 and to a mention in Gartner’s 2026 Hype Cycle for Platform Engineering as a sample vendor in software supply chain security. The product, Valtman says, was shaped by three recurring pieces of customer feedback: that where findings land matters more than what they find, with issues surfaced at the source-code-management layer getting resolved because developers handle them inside code they are already changing; that fine-grained policy control over finding metadata and staged rollout was a primary selection criterion; and, most recently, that hybrid code from humans and AI agents is now being deployed faster than security teams can manage.
The choice to preview the product at OWASP’s European gathering was deliberate. “These are our people,” Valtman said of the AppSec practitioners and CISOs who fill the conference, a community Arnica has courted through years of top-tier sponsorship. He is also candid about where he sees genuine innovation, and where he does not. The most interesting work, he observed, is happening inside enterprises building their own tooling, such as custom dashboards. A handful of companies had tried building their own AI scanners; all of them, he said, walked away, deterred by the operational risk and the cost of maintaining such a system in-house.
That last detail doubles as Valtman’s thesis. As AI writes an ever-larger share of enterprise software, the differentiator will not be who can generate code fastest. It will be who can prove, when a regulator or an auditor asks, that someone was accountable for it.
Disclaimer: This article contains sponsored marketing content. It is intended for promotional purposes and should not be considered as an endorsement or recommendation by our website. Readers are encouraged to conduct their own research and exercise their own judgment before making any decisions based on the information provided in this article.
