The Governance Gap

AI production deployment — especially in R&D, with agent-based coding — appears to be moving faster than the IT governance structures needed to manage it. Securing AI products appears to be lagging behind AI adoption. While IT guardrails are often a blend of automated and manual controls, our data suggests a larger reliance on manual controls around AI products. As a result, we see AI-related security incidents happening at scale. This gap is not theoretical, and it appears to be widening.

The governance gap is not a risk on the horizon — it is a production reality. More than half of B2B software technical decision makers report experiencing an AI-related incident in the past year. The most common is hallucinations causing business errors, at 40% — statistically equal to the 44% who cite hallucinations as their top barrier to scaling agentic AI.

While the impact of some of these incidents is only internal, we believe the potential for external incidents such as data leakage, jailbreak attacks, and compliance exposure sits at levels unacceptable for the vast majority of technical decision makers.

55% of all technical decision makers say AI adoption has moved ahead of their security capabilities — but the organizations furthest along in AI deployment are the most aware of the security lag their own velocity has created. We view this not as a comment on security at Runner-tier organizations, but as evidence that deployment speed has outpaced security program adaptation.

While the data shows most technical decision makers using around three AI guardrails on average, it also suggests a guardrail automation gap may exist between manual approaches and automated ones. Human approval gates are the most common single guardrail (75%), and the most formalized automated enforcement is the least adopted (policy-as-code, 29%) — suggesting teams lean on human review and lighter-weight automated controls rather than codified runtime policy.

The agentic threat landscape has a clear center of gravity among those surveyed. Excessive agency — over-permissioned agents with read, write, or deploy access that create a large blast radius when compromised — is ranked #1 by 32% of technical decision makers. Prompt injection — inputs that redirect agent behavior — is ranked #1 by 28%. Together they account for 60% of all #1 rankings; no other category reaches 15%. Both are characteristically agentic: they exist because agents take autonomous actions in response to inputs. These are new threat categories directly named as a result of AI.

As this benchmark series continues, we will track the evolution of cybersecurity and its ability to reduce AI threat-vector concerns and reduce incident rates.

AI security appears to be entering into consideration earlier in the AI development lifecycle — 63% of technical decision makers now integrate security at the requirements or early-development phase, and Runners lead at 73% early integration versus 56% of Walkers. But 31% still integrate security late in development or only after deployment, and 5% report no formal AI security integration at all. Late-stage security integration may be linked to the higher AI incident rates noted in Finding 1.

The governance gap is not only a security-tooling story — it is reflected in organization design. Roughly a third of surveyed B2B companies have no formal AI governance structure at all, and the shortfall is not evenly distributed. Ad hoc is the single most common operating model on the go-to-market side (33%), while technical decision makers most often report a hybrid structure and are less likely to be operating with no structure at all (27%). The starkest contrast is the centralized Centre of Excellence: 25% among technical decision makers against 16% in GTM.

We believe this is a gap that needs to be filled quickly and adequately. It is not just a security risk; it is a general risk to your AI strategy overall, across all dimensions — productivity, ROI, and costs — along with security concerns. And on the evidence here, the commercial organization is roughly a maturity rung behind the technical one in closing it.