In July, Hugging Face disclosed that an autonomous AI agent had spent days operating inside its production environment without human involvement. What seemed like an isolated incident is now emerging as a trend.
By September, reports described agents bypassing controls on an Australian government health portal, uploading dozens of user images to third-party sites and taking over a dormant wiki. Notably, several of these incidents went undetected for months before being discovered and disclosed.
In a Cloud Security Alliance survey of more than 400 IT and security professionals, 65% reported experiencing an AI agent-related incident in the past 12 months, and 61% reported data exposure as an impact.
As agents and models become more capable, the lesson is not that agents are too dangerous to deploy. It is that any single control can fail. Governance and controls must be layered and independent, so that one gap does not become an incident.
The industry is moving in the right direction. In late September, NVIDIA launched its Open Agent Safety Platform, with more than 100 organizations working with its technologies. NVIDIA’s OpenShell provides an open-source secure runtime that isolates agents and enforces access policies outside the agent process. For supported BlueField-4 architectures, the Sentry reference design adds an independent monitoring and enforcement layer that can quarantine an agent attempting to exceed its boundaries.
However, no single layer is enough. The agents in the Hugging Face incident exposed a flaw in a path they were permitted to use, which is the kind of gap that one layer alone may not catch. Layers that fail independently keep one gap from becoming a breach.
Three layers of control
The readiness gap
Governance is often treated as a bottleneck. In regulated industries, it enables enterprises to safely scale agentic AI.
In Kyndryl’s People Readiness Report, a survey of 1,100 business and technology leaders found that just 23% of organizations considered their governance and compliance ready to support AI adoption. Only 41% were confident in the guardrails their organization had put in place.
This readiness gap may widen as agents take on more consequential work. Closing it will require building layers that hold and protect against any one failing.
Where to start
Organizations deploying agents today can take four practical steps.
01
Inventory agents
In the Cloud Security Alliance survey, 82% of respondents reported discovering previously unknown agents in the past year. Enterprises cannot govern what they have not inventoried. Kyndryl’s own research found that just 27% of leaders were fully confident that their organizations maintained a central registry and monitored all AI systems operating in their environments.
02
Put highest-risk rules into code first
Start with actions that are hard to reverse: deleting data, moving money, changing production systems. Route every agent action that could trigger them through an enforcement point.
03
Add layers that can operate independently
Pair business policy with runtime isolation, hardened infrastructure and out-of-band monitoring — layers that depend neither on the agent nor on each other.
04
Include evidence from day one
Log each decision with the rule that allowed it and the person who approved it in a form regulators accept.
Agents will keep getting more capable, and new ways for them to fail will follow. The goal for leaders is to build governance and security into the runtime so that when the unexpected happens, independent layers of control stop it from becoming a business disruption.