Why agent detection gaps are your biggest AI risk (and how to close them)

In this week's headlines I flagged Microsoft's Agent Governance Toolkit, fresh Cequence/EMA research on detection failures, and the Forbes reminder that agents are already running inside your stack whether IT signed off or not. Same story: production moved faster than security operations.

The scary number is not "65% had an AI incident." It is that only about one third of organizations can contain an out-of-scope agent action in minutes with automation. Everyone else is on hours, manual steps, or finding out during a scheduled review. That is not a model problem. It is an operations problem you can fix in a week if you are honest about gaps.

What the research actually says

Enterprise Management Associates surveyed enterprises deploying agentic AI and found widespread production use paired with weak controls:

• 65% report at least one AI-related security incident.
• 32% can detect and contain out-of-scope agent activity within minutes using automated mechanisms.
• 55% need hours and manual intervention.
• 46% cannot easily produce a complete 30-day audit trail for a specific agent.
• 47% cannot reliably inventory all deployed agents.

Separate vendor surveys echo the visibility gap: a large share of agents run with little or no security oversight, while executives still believe written policies are enough. Policy without runtime enforcement is paperwork.

Why logs alone fail

Traditional security stacks were built for humans clicking in browsers and services calling APIs with stable identities. Agents add non-deterministic tool chains: read CRM, draft email, post to billing, retry on error. SIEM rules written for "user X accessed file Y" break when the actor is an agent with delegated OAuth scopes and ten intermediate steps.

You need three capabilities that logs do not provide by themselves:

• Inventory: know every agent, owner, and tool surface.
• In-path enforcement: evaluate each tool call before it runs.
• Containment: revoke credentials and halt workflows in minutes, not during the quarterly review.

Microsoft's open toolkit and commercial gateways differ in packaging, but they agree on the architecture: governance belongs between the agent and the system of record.

A one-week sprint for a mid-size operator

Day 1: Inventory (no new tools required)

Export a spreadsheet with columns: agent name, business owner, technical owner, data classes touched, tools/MCP servers, environment (prod/pilot), last change date. Include browser copilots, SaaS agents, Zapier/Make flows, and internal scripts that call LLM APIs. If you cannot fill a row, mark it "unknown" and assign someone to resolve it by Friday.

Day 2: Define "out of scope"

Write five concrete examples for your business: agent emails a customer without approval, agent reads payroll, agent creates a vendor payment, agent exports more than N records, agent calls a tool not on the allowlist. Share the list with engineering and support. Ambiguous scope is why incidents take hours.

Day 3: Containment drill

Pick one non-critical production agent. Simulate compromise: rotate its API key, disable its MCP server, or flip a feature flag. Time how long until all tool access stops. If it takes more than 15 minutes, you have your board slide.

Day 4: Audit trail sample

Ask for a report: every action agent X took in the last 7 days, with timestamps, inputs redacted, and approver if applicable. If vendors cannot produce it, negotiate before you expand scope.

Day 5: Decommission policy

Document agent offboarding: revoke OAuth, delete secrets, remove schedules, archive logs, notify owner. Apply it to one retired pilot this week. EMA lists decommissioning as a first-class discipline for good reason.

What to buy vs what to build

Build if you have one or two internal agents and a strong platform team: use open governance kits, strict tool allowlists, and central logging with immutable storage.

Buy when agents touch ERP, CRM, or payments across vendors. You want a gateway that scores risk per tool call, routes high-risk actions to human approval, and exports evidence for SOC 2 and EU AI Act reviews without a custom integration per system.

Either way, procurement should ask: show me inventory, show me block in under 15 minutes, show me a 30-day trail. If the demo is only a chat UI, keep shopping.

What to ignore

• Benchmark scores as proof of security.
• "We have a responsible AI policy" without runtime controls.
• Discovery tools that never connect to enforcement.
• Perfect inventory before you start; imperfect inventory beats denial.

Bottom line

Agentic AI rewards speed. Incidents reward preparation. Close the detection gap with inventory, in-path policy, and a rehearsed kill switch before you scale the next workflow. Your CFO and your insurer will ask for evidence sooner than you think. Pair detection with human approval queues and this week's governance headlines on kill switches and board KPIs.

Questions on agent governance for your stack? Reach out and we will point you at the right pattern.

Matt Potter · Swift Media