88% Of Organisations Had An AI Agent Security Incident Last Year - And Half Of Employees Now Use Agents Daily. Welcome To The Age Of 'Agent-Ops', Where Running AI Safely Is The New Enterprise Skill

88.4% - Organisations that experienced at least one AI-agent-related security incident in the past year (AvePoint 2026 State of AI)  ·  46.9% - Employees now using AI agents weekly or daily - adoption has outrun the discipline to run agents safely  ·  Agent-ops - The emerging operational practice of deploying, monitoring, securing and governing fleets of agents in production  ·  Mainstream - Decision governance for AI agents has moved from niche concern to a mainstream enterprise priority (Gartner)

Here is the statistic that should reframe every UK enterprise's AI conversation this quarter: 88.4% of organisations experienced at least one AI-agent-related security incident in the past year, according to AvePoint's 2026 State of AI report - even as 46.9% of employees now use AI agents weekly or daily. Read those two numbers together and the picture is unmistakable: AI agents have spread through the workforce far faster than the discipline to run them safely. Nearly half of employees now lean on agents routinely, and nearly nine in ten organisations have already been bitten at least once.

That gap - between how fast agents have been adopted and how slowly the operational discipline to control them has caught up - is the defining enterprise-AI problem of the second half of 2026. And it has a name emerging to describe the answer: 'agent-ops', the operational practice of deploying, monitoring, securing and governing fleets of AI agents in production. As Gartner's 2026 analytics trends put it, the priority is shifting from simply deploying models to governing, securing and operationalising AI decisions at scale. As an AI Agency London that runs agents in UK production, we think agent-ops is about to become as fundamental an enterprise capability as IT operations or cybersecurity - and this article is the honest read on why, and how to build it.

The good news is that the infrastructure to do this properly is arriving fast. NVIDIA has shipped an Agent Toolkit specifically for safely managing autonomous agents in production, HPE and NVIDIA are building AI Factory infrastructure with dedicated agent-orchestration hardware, and a wave of funding is flowing to agentic platforms built for regulated sectors - 8090 Labs just raised $135 million, led by Salesforce Ventures, for an agentic system aimed at healthcare and aerospace. The tools to run agents safely exist. The question for UK enterprises is whether they build the discipline to use them before they become another entry in the 88%.

Why The 88% Happened: Adoption Outran Control

The agent security reckoning is not a story of exotic attacks - it is a story of ordinary loss of control. When nearly half of employees adopt agents routinely, much of it happens informally: staff connect agents to company systems, grant them access to data, and let them take actions, often without central oversight. This is 'shadow AI' - the agentic successor to shadow IT - and it is where most of the 88% of incidents originate. An agent given broad access by a well-meaning employee can expose data, take wrong actions, or become a security hole, not through malice but through the absence of the controls that would normally govern such access.

The core problem is that agents are powerful in a way that ordinary software tools are not. A traditional app does a defined thing; an agent can plan, decide and act across systems, which is exactly what makes it valuable and exactly what makes ungoverned agents risky. The organisations in the 88% mostly did not make a dramatic mistake - they simply let agent adoption run ahead of the discipline to govern it. The fix is not to slow adoption, which is neither possible nor desirable, but to build the agent-ops capability that lets agents be used widely and safely at the same time.

The Five Pillars Of Agent-Ops

1. Visibility: know every agent operating across the business, including the informal ones staff have connected to company systems. You cannot secure an agent fleet you cannot see, and shadow agents are where most incidents start.
2. Access control: govern what each agent is allowed to do and reach, on a least-privilege basis - an agent should have access to exactly what its job needs and no more.
3. Monitoring: watch what agents are actually doing in production in real time, so anomalous or risky behaviour is caught quickly rather than discovered after an incident.
4. Audit: keep a complete, tamper-evident record of every consequential agent action and the data behind it - essential for security, for trust, and for the auditability regulators increasingly require.
5. Response: have a plan and the controls to intervene fast when an agent misbehaves - pause it, revoke access, roll back - so a problem is contained rather than compounded.

Agent-Ops Is What Lets You Scale Agents, Not What Slows Them

It would be easy to read the 88% figure as a reason to be cautious about agents. That is the wrong lesson. Agents are delivering real value - which is exactly why half of employees have adopted them without waiting for permission. The right lesson is that the businesses which build agent-ops can safely let agents do far more, because they can see, control, monitor and trust them. A business with strong agent-ops can confidently expand its agent fleet into higher-value, more sensitive work; a business without it has to either restrict agents severely or accept being in the 88%. Agent-ops is not the brake on agentic AI - it is what makes scaling it safely possible.

For UK enterprises in regulated sectors, this is doubly true. The same agent-ops disciplines that prevent security incidents also produce the visibility, control and auditability that FCA, ICO, MHRA and SRA oversight increasingly expect - and that the EU AI Act requires for high-risk systems. Building agent-ops well is therefore not just risk reduction; it is the foundation that lets regulated UK businesses deploy agents in exactly the high-value areas their less-disciplined competitors are too exposed to touch.

The 90-Day Agent-Ops Playbook For UK Enterprises

1. Days 1-20: Run an agent-discovery exercise across the business - find every AI agent connected to company systems and data, including informal ones spun up by staff. Map what each can access and do.
2. Days 21-40: Apply least-privilege access to every agent - restrict each to exactly the systems and data its task requires - and shut down or bring under management any shadow agents that pose real risk.
3. Days 41-60: Stand up monitoring and audit logging so every consequential agent action is visible in real time and recorded with its data lineage. Prioritise agents touching personal or regulated data.
4. Days 61-80: Build an incident-response capability for agents - the ability to pause, revoke or roll back an agent fast - and rehearse it. Map your controls to FCA, ICO, MHRA, SRA and EU AI Act expectations.
5. Days 81-90: Appoint a named owner for agent-ops and make it a standing function, not a project, so the discipline keeps pace as your agent fleet grows across the business.

Sources

1. AvePoint - '2026 State of AI' report (46.9% of employees use AI agents weekly/daily; 88.4% of organisations had at least one agent-related security incident)
2. Gartner - 2026 Data & Analytics trends: decision governance for AI agents moving from niche to mainstream
3. NVIDIA - Agent Toolkit for safely managing autonomous agents in production (via HPE AI Factory expansion)
4. HPE - AI Factory portfolio expansion with NVIDIA (Vera CPU for agent orchestration)
5. TechCrunch / BuildFastWithAI - '8090 Labs closes $135M Series A led by Salesforce Ventures for agentic Software Factory' (July 2026)
6. BraivIQ - Batch 27 AI Governance & Guardrails article (internal reference)