Share with your CISO
Enterprise AI deployment has outpaced the security programs meant to govern it, and a new DigiCert survey of 1,001 IT and cybersecurity decision-makers puts numbers to the gap. Seventy-eight percent of organizations reported AI-related security incidents or identified vulnerabilities, while 75% deployed four or more AI-powered systems in just six months. Nearly half lack centralized visibility into those systems. Separate Gartner research found 62% of organizations hit by deepfake-driven social engineering attacks in the past year, with 32% facing adversarial prompting and 29% seeing attacks on AI application infrastructure directly.
What this means for your business
The uncomfortable truth buried in these numbers is that governance theater is already widespread. Ninety percent of organizations discuss AI governance at the board level, but only about half have formal programs, and 47% cannot trace an AI decision back to the model or data that produced it. If you can’t answer “which system did this, what did it touch, and who owns the response,” you don’t have an AI security posture. You have a slide deck. The organizations most exposed are those that treated pilot deployment as a security checkpoint rather than the beginning of one.
The traceability gap is the sharpest operational problem here, and it deserves more attention than it typically gets in governance discussions. Traceability means the ability to reconstruct, after the fact, which model produced an output, what data it accessed, and through what workflow it acted. Without it, incident response becomes forensic archaeology. That matters doubly for AI agents, which increasingly hold credentials, call APIs, and trigger downstream processes autonomously. An agent that gets compromised mid-workflow doesn’t leave the same fingerprints a human user does, and existing SIEM and EDR tooling wasn’t designed to monitor non-human actors moving through enterprise systems at machine speed.
Deepfake-driven social engineering is the threat that should force a budget conversation CISOs haven’t been able to win before. Voice cloning and synthetic video have crossed the threshold where older phishing awareness training, built around suspicious links and grammatical errors, fails on first contact. Sixty-two percent of Gartner’s sample already experienced this. The question for any CISO defending a flat security budget is whether their organization’s incident response plan names a specific owner and a specific stopping mechanism for an AI agent that has been manipulated or compromised. If that gap exists, the DigiCert data suggests it’s not hypothetical exposure. It’s a matter of when.
Concept deep-dive: Adversarial prompting
Adversarial prompting is the practice of crafting inputs to an AI model specifically designed to make it ignore its instructions, leak data it shouldn’t surface, or take actions outside its intended scope. Think of it as the AI equivalent of SQL injection, where the attack lives in the data, not the network perimeter. Thirty-two percent of Gartner’s respondents experienced this. For security teams, it means the model itself is an attack surface, not just the infrastructure surrounding it.
Based on reporting from AI Governance Gap Widens as Enterprises Report Security Incidents, originally published 2026-07-08 03:00:00.
