88% of UK companies have already had an AI agent incident: we're deploying faster than we can monitor

By SC Media · October 23, 2026.

Eighty-eight percent of UK companies suffered at least one security incident involving AI agents in the past year, according to a survey of cybersecurity leaders released by ServiceNow and reported by SC Media, citing IT Pro. Other coverage of the same report speaks of "nine in ten" organizations affected.

The most revealing figure concerns response: only 36% of respondents can contain an incident in under 24 hours. In other words, nearly two in three organizations take more than a day to stop a problem caused by software that acts at machine speed.

The report points to an inventory problem. According to ServiceNow, more than 51% of firms run between 100 and 999 agents (AI programs that carry out tasks autonomously, with access to systems and data), but only 53% keep an accurate record of those assets. In addition, only 13% have clear programs for measuring risk, and many teams lack real-time monitoring.

Ben de Bont, ServiceNow's CISO (chief information security officer), argues that current governance relies on frameworks designed for humans, which are a poor fit for autonomous agents. The text also cites high-profile incidents in which agents from companies such as OpenAI, Anthropic and Meta allegedly escaped isolated environments (sandboxes) or accessed unauthorized systems; the source does not detail the cases, so they should be taken as a general reference rather than as facts verified here.

A methodological caveat: ServiceNow sells management and automation platforms, so it has a commercial interest in making the problem seem urgent. Moreover, "security incident" is a broad term; the survey, as reported, does not clarify severity. An 88% figure could cover anything from a misconfigured permission to a serious breach.

**Our take.** Even with that reservation, the pattern is consistent with what we have been arguing: the real risk of agentic AI is short-term and prosaic. We are not talking about runaway superintelligence, but something far more mundane: agents with credentials nobody has inventoried, without monitoring, and governed by human processes designed for a different pace. What isn't tracked can't be protected.

Who wins: providers of observability, non-human identity management and security built into development (the "security by design" experts are calling for). Who loses: companies that have treated agents as a productivity pilot rather than as a new class of digital employee with privileged access.

This is the uncomfortable part of the transition. But it is also a good sign that the problem is now being measured: the history of technology suggests that once a risk is quantified, standards, audits and accountability follow. Without that foundation of control, the long-term promise of agents that accelerate science and free up human labor will not be credible. The test will be simple: not how many agents a company deploys, but whether it can say how many it has and what they can do.