August 2026 · 6 min read
From 15 to 150,000 Agents: The Production Governance Gap
Key Definitions
Standing Credentials Broad, never-expiring access granted to an agent — an agent that only needs to read one file is given permanent read access to the whole folder; an agent that needs to send one email is given the entire inbox.
Just-in-Time (JIT) Permissions Scoped permissions auto-provisioned for the specific task at hand and revoked the moment it completes, replacing standing credentials so each agent holds only the minimum access it needs at the moment it acts.
Somewhere in a mid-sized financial services company right now, an AI agent is reading a customer's account history, pulling data from three internal systems, and deciding what to do next — no human in the loop and, in nearly half of the cases, without anyone watching. This is not hypothetical. It is the baseline picture painted by Okta's "AI Agents at Work 2026" report: agents have quietly moved from demo projects to the most active — and least governed — actors in production systems.
Agents Are Exploding Faster Than Governance
A few numbers frame the scale. According to Okta's "AI Agents at Work 2026" report, 88% of organizations have already experienced suspected or confirmed AI agent security incidents, and 90% have no reliable way to govern what their agents actually do in production. Meanwhile, the average Fortune 500 company is projected to run 150,000 AI agents by 2028 — up from fewer than 15 just last year.
The deeper problem is the governance model itself. Omdia research finds that 96% of organizations run AI agents on governance models that were not designed for them. Traditional security was built for humans: log in, get the access you need, get the work done, log out. Agents are software acting autonomously inside sensitive systems — they initiate actions and decide which tools to call on their own, almost never passing through traditional security workflows.
The Core Problem: Standing Credentials That Never Expire
Companies first treated agents like ordinary software tools — they got credentials, the same as any other system: a service account here, an API key there. It seemed fine at the time. It no longer is. Most enterprises give their agents broad, standing credentials that never expire. An agent that needs to read a file gets permanent read access to the entire folder. An agent that needs to send one email gets access to the entire inbox.
When an agent acts unexpectedly or gets compromised, the blast radius is maximized. Security practitioners now call this the non-human identity problem: when an agent is compromised, there is often no registry showing what the agent was, what it had access to, or a way to shut it down. An attacker gains not a single-use key, but the full set of the agent's standing access.
Why Traditional Identity Management Fails for Agents
Traditional identity management revolves around humans: people log in, they log out, and their identity lifecycle is clear. Agents do not follow that pattern. Hush Security co-founder and CEO Micha Rave puts it plainly: "Every company already knows how to manage identity for its people and its applications. But now software acts autonomously, on its own initiative, inside your most sensitive systems. AI agents need strict identity, not just API keys."
That judgment has spawned a rapidly forming governance market. On July 28, 2026, Hush Security closed a $30 million Series A round, led by Battery Ventures and YL Ventures, with Akamai Technologies joining as a strategic investor, bringing total funding to $41 million. The company, which emerged from stealth in September 2025, was selected for the 2026 CrowdStrike/AWS/NVIDIA Cybersecurity Startup Accelerator. Capital's choice is itself a signal: the agent governance gap is not theoretical — it is a validated, real market.
Governing Agents in Production: Four Building Blocks
Whatever vendor's platform you choose, governing production agents converges on the same architecture. These four building blocks are all required:
Central registry
Enroll every agent in a central registry, identify agents running in your environment, and map the MCPs, tools, and resources they reach and use. Agents you cannot see cannot be governed.
Just-in-time scoped permissions
Replace standing credentials with JIT scoped permissions: auto-provisioned for the specific task being performed and revoked when the task completes, eliminating never-expiring broad access.
Runtime detection and audit
Log every action, detect at runtime how permissions are actually being used versus how they were granted, and maintain a complete audit trail for each agent interaction to support investigations and compliance.
A centralized kill switch
The ability to shut down any agent immediately. When you cannot turn off a compromised agent, all you have is the right to watch the damage it can cause.
OOMeta's View
The explosion of agents makes "invisible and ungovernable" the #1 structural risk in enterprise AI. This has nothing to do with how many agents you deployed last month, and everything to do with whether you can answer three questions: Is every one of my agents in some queryable registry? Can I state precisely what permissions each agent currently holds and whether they are being used as granted? Can I shut down an anomalous agent at any moment? Treat agents as first-class security principals — give them identity, scope, audit, and a kill switch instead of a never-expiring API key — and you won't be quietly taken over by the automation you deployed.
References: SecurityWeek, "Hush Security Raises $30 Million for AI Agent Governance", 2026-07-28, https://www.securityweek.com/hush-security-raises-30-million-for-ai-agent-governance/；Enterprise DNA (via SecurityWeek) citing Okta's "AI Agents at Work 2026" report and Omdia research, "AI Agent Security Incidents Hit 88% of Enterprises", 2026-08-03, https://enterprisedna.co/resources/news/hush-security-30m-ai-agent-governance-enterprise-2026
Frequently Asked Questions
Why is the agent count growing so uncontrollably?+
Fortune 500 companies are projected to grow from fewer than 15 agents last year to 150,000 by 2028. Deployment has incentive — teams want automation, budgets exist, checklists get executed — but there is almost no central registration or retirement, so volume explodes while governance does not keep up.
What does '90% can't govern production behavior' mean?+
It means that once an agent acts unexpectedly or is compromised, there is often no registry showing what the agent was, what it had access to, or how to shut it down. Governance is reactive, incomplete, or absent entirely.
What is wrong with standing credentials?+
An agent that needs to read one file is permanently granted read access to the whole folder; an agent that needs to send one email gets the entire inbox. The credentials never expire, so a compromised agent hands an attacker its entire standing access.
Why does traditional identity management fail for agents?+
Traditional identity is built for humans — log in, get access, log out. Agents are software acting autonomously inside sensitive systems, initiating actions on their own. As Hush Security's CEO puts it, AI agents need strict identity, not just API keys.
How should enterprises govern agents in production?+
Enroll every agent in a central registry, replace standing credentials with just-in-time scoped permissions, detect at runtime how permissions are actually used versus how they were granted, keep a full audit trail of every action, and provide a centralized kill switch to shut down any agent immediately.
Related Articles
AI Agents Retire Too: The Unmanaged Decommissioning Gap
Deploying agents is heavily documented; retiring them is not. Un-retired agents leave live credentials, standing access, and unattributable spend — the most preventable dark matter risk.
Governance Decay: Context Compaction Erases Agent Safety
A June 2026 paper (arXiv:2606.22528) names enterprise AI's quietest failure mode — governance decay — where context compaction drops "outdated" safety rules and the same agent later acts without a visible signal.
The Agent Identity Ownership Vacuum: Who Answers for Autonomous Systems
A CSA/Strata survey of 285 security professionals finds only 23% have a formal agent identity management strategy, with fragmented ownership and identity infrastructure that cannot support autonomy in production.
88% of Enterprise AI Agents Fail Security Tests: The Control Gap Widens
Enterprise agent adoption is outpacing the security control that follows it, making the control gap the defining enterprise AI security problem of 2026.