ServiceNow says that ungoverned agentic AI exposure is compounding inside the enterprise, based on new global research of senior security leaders. According to a survey for the ‘Deployed but ungoverned: the state of agentic security’ report, 90% of organizations around the world have experienced an agentic AI security incident in the past year.
The survey found a severe gap between AI deployment and security, with many security leaders stating that they have established agentic security postures, but only 13% agreeing that they have defensible programmes when measured against five key domains: attack surface, accountability, security by design, visibility, and audit readiness.
Agentic AI governance
Most agentic AI is currently governed by legacy frameworks and human-driven systems, with less than half of organizations using governance frameworks that cover key elements of AI agent deployments. Just 44% of security teams globally design security into an agent’s development before it’s built, which means that, for the majority, permissions and guardrails are being implemented on an ad hoc, after-the-fact basis.
Key AI governance findings:
- 49% of organizations now run between 100 and 999 AI agents, and 35% run more than 1,000.
- 57% of security teams claim to have a full, accurate record of every AI agent in production and have processes to track changes.
- 29% of security teams found AI agents running in production that had never been registered or known.
- 76% of organizations say they maintain a complete, current map of every agent’s access, but only 19% can see more than 90% of their agents in real time.
- 66% cannot act automatically when an agent behaves anomalously, which means a human has to intervene before anything changes.
Methodology
‘Deployed but ungoverned: the state of agentic security’ was conducted by ServiceNow in conjunction with Phronesis Partners. The study surveyed more than 3,700 senior security and technology decision-makers, VP level and above, across 20 countries. Respondents represent mid-sized and large organizations, with 97% reporting annual revenue of $250M or more.
Resilience Forward comment
Challenge questions for boards on agentic AI
The starkest warning in this research is the speed mismatch: autonomous AI agents execute at machine speed, yet two-thirds of organizations rely on manual human intervention when an agent behaves anomalously. Combined with widespread ‘shadow’ agents and a critical lack of real-time visibility, boards cannot simply accept leadership assurances on AI governance. To provide effective governance, board members need to challenge executive teams. Three potential questions are:
- How do we verify that our AI agent inventory is based on real-time discovery rather than self-reported registers?
- What automated containment controls exist to halt anomalous agent behaviour before it breaches operational resilience impact tolerances?
- Are security and least-privilege access controls embedded before agents enter production, or applied ad hoc?
