Four in five firms have agents they cannot see. The fix on sale is the inventory

The agents you do not know about

The numbers in this week's agent governance announcements are stark, and they come from vendor-cited surveys, so they should be read as a market's self-portrait rather than independent measurement. Fiddler cites a survey finding that 82% of organisations had unknown AI agents running in their infrastructure as of January 2026, and that 65% experienced an AI agent-related incident in the 12 months before the survey. Data exposure was a business impact for 61% of AI agent-related incidents, according to the same survey, and 75% of organisations report unauthorised use of AI coding assistants. m:…0592

Linx cites a survey with the same shape. It says 77% of surveyed enterprises have AI agents in production, but only 20% of organisations using or exploring agents report having a complete agent inventory. Half of organisations report employee access to public software-as-a-service generative AI agent platforms, according to Fiddler's cited survey. The picture is consistent across the two vendors: agents are in production almost everywhere, and the list of what is running is incomplete almost everywhere. m:…5830 m:…0592

The threat the surveys point to is not a rogue model. It is a legitimate credential doing the wrong thing. Linx's cited survey says 51% of CISOs identify compromised or hijacked agents as a top AI-related threat. Only 34% of organisations use dedicated per-agent identities with scoped permissions, while 26% use human credentials passed to agents and 15% use shared service accounts. The remaining gap is the attack surface: an agent running on a human's credentials inherits a human's access. m:…5830

Inventory first, then control

Fiddler's prescription starts with the list. The company says enterprises should treat every AI agent as a non-human identity with a named human owner, scoped credentials, and a documented purpose, and that an AI agent inventory should be populated from telemetry, gateway traffic, and identity data. Its platform builds agent records from production telemetry, according to the company. The regulatory hook is real: the EU AI Act requires high-risk AI systems to allow automatic recording of events over the system's lifetime, and NIST AI RMF 1.0 GOVERN 1.6 requires mechanisms to inventory AI systems, according to Fiddler. m:…0592

Fiddler also draws a line buyers often blur. Governance defines and enforces what an agent is permitted to do before execution, while observability records and explains what an agent did during and after execution. Regulatory frameworks like the EU AI Act require both control enforcement and evidence recording, the company says, and handoffs between compliance and platform engineering are where controls tend to fail. Policy decisions must be written into telemetry traces so auditors see the control next to the action. Traditional application performance monitoring is insufficient for agents, Fiddler argues, because agents can produce different plans from the same input. m:…8175

Banks get a dedicated tool

Uptiq launched GovernAI this week, a central governance tool for financial institutions. The company says it manages a central registry of agents across the enterprise, grounds every AI agent in the institution's truth, enforces role-based permissions for rollouts, and mandates human approval at defined steps. It displays the exact source document behind every figure in an AI-prepared credit memo, enables independent evaluation of agent performance, and verifies through regression testing that policy updates do not alter agent outputs before going live. GovernAI is available immediately for all Uptiq customers at no additional cost, according to the company. m:…5130

The buying intentions in Linx's survey suggest the market is moving. It says 55% of CISOs plan to evaluate agent identity, authorisation, and access governance tools in the next 12 months, 49% plan to evaluate AI-powered risk scoring and behaviour analytics, and 48% plan to evaluate runtime guardrails for production agents. At the same time, 88% of surveyed organisations permit AI to execute some security response actions without human approval, though 71% limit autonomous security response to low-risk actions. The governance gap and the autonomy are growing together. That is the risk the inventory is meant to close, and the reason a registry is being sold before a control plane. m:…5830