AI-Assisted Development Security: What XBOW’s CISO Gets Right

The News

XBOW, an AI-powered offensive security company, appeared on the AppDevANGLE podcast to discuss the security implications of AI-assisted software development. Nico Waisman, XBOW’s CISO and head of offensive product capabilities, drew on his background at Lyft, Semmle, and GitHub Advanced Security to explain why the speed of AI code generation is creating a new category of technical and security debt. The conversation covered guardrail architecture, the build-versus-buy calculus for autonomous AI tooling, and the governance structures organizations need before AI-generated code becomes ungovernable.

Analyst Take

The debt is already accumulating, and most teams don’t see it yet

AI-native technical debt is not a future problem. It’s accumulating right now, sprint by sprint, in organizations that treat AI-assisted development as a productivity win without building the feedback loops to validate what’s actually being shipped. The conventional security review cadence, quarterly pentests, manual code reviews, SLA-driven CVE triage, was already straining under modern DevOps velocity. AI code generation doesn’t just strain that model further; it makes it structurally irrelevant.

What makes this especially critical is the volume of code involved. According to ECI Research’s Google GovTech Survey Results, 49.6% of respondents selected “26% to 50%” when asked what percentage of their organization’s code they estimate will be assisted or generated by AI within the next 12 months. That’s not a pilot. That’s a substantial share of production logic being written by systems that carry no organizational context, no memory of past compliance decisions, and no accountability when something breaks. Security and engineering leadership need to internalize that number before they can have an honest conversation about risk posture.

The “paved road” model is the right answer, and most organizations aren’t building it

Waisman’s prescription is grounded in a philosophy that engineering-led companies like Netflix and Lyft operationalized years before AI entered the picture: make the secure path the fast path. He described the concept of building a “golden path,” where doing things correctly requires no friction while deviating from approved patterns slows you down. This is exactly the kind of architectural thinking that scales alongside AI code generation. What doesn’t scale is the alternative: siloed security reviews bolted on at the end of a sprint cycle.

The data supports how far most organizations are from this model. According to ECI Research’s Google GovTech Survey Results, 31.0% of respondents said that “Security is treated as a final gate, causing significant rework” when asked how often security compliance mandates force developer teams to rewrite or discard completed application code. Nearly half said this happens “occasionally, requiring mid-sprint corrections.” When you layer AI-generated code volumes on top of a late-gate security model, the rework cost compounds quickly. The paved road architecture Waisman describes isn’t aspirational; it’s the only model that remains economically viable at AI-assisted scale.

Build versus buy: the POC trap

One of the most practically useful observations in this conversation is Waisman’s account of the build-versus-buy cycle his customers keep repeating. The pattern is consistent: a team decides to build an internal autonomous security or AI tooling capability, gets a prototype working in weeks, then returns to XBOW two or three months later having discovered that safety, reliability, and cost containment at production scale are far harder problems than the POC suggested. LLM-based systems introduce indeterminism that requires continuous benchmarking. Model updates can silently shift application behavior. And without the right guardrails, inference costs can exceed what a commercial solution would have cost entirely.

This is a decision that ITDMs need to own explicitly, not delegate to engineering. The appeal of internal control is real, and there are legitimate reasons to build custom tooling in some contexts. But the total cost of ownership calculation needs to include the ongoing engineering overhead of maintaining benchmarks as models evolve, building safety harnesses that can’t simply be prompted into existence, and staffing the governance function that ensures the system behaves as intended over a two- to three-year horizon. ECI Research’s Google GovTech Survey Results found that 47.2% of respondents selected “Developer velocity and ease of integration” as the greatest weight in their final technical selection process. Speed is driving build decisions that will be expensive to unwind.

Looking Ahead

The governance question Waisman raises at the end of the conversation is the one that will define success in enterprise AI security over the next 18 months. Organizations that treat AI governance as a centralized function owned by a single team will repeat the mistakes of first-generation cloud adoption, where sprawl and shadow IT emerged precisely because central control didn’t move at the speed engineering teams required. The model that works, as Waisman describes it, distributes ownership across every team building software while holding the strategy at the executive level. That’s a cultural and organizational change, not a tooling purchase, and it takes longer than most roadmaps currently account for.

XBOW’s positioning as an AI-native offensive security company with nearly three years of production deployments and a customer base in the hundreds gives it meaningful data advantages over newer entrants making similar claims. The open-source benchmarks the company released, already solved at 85% by leading models before XBOW had to retire and replace them, illustrate both the pace of model capability growth and the difficulty of maintaining meaningful evaluation frameworks over time. Expect the benchmark arms race to accelerate through 2027, and expect the organizations that invested early in continuous testing pipelines to show measurably better security posture outcomes than those still relying on periodic assessments.