The State of Autonomous
Agent Governance
We asked 300 CISOs and CSOs how they run, secure, and trust the agents already inside their companies. Three in four are very confident they know what is connected. Half found an agent this year with access it should never have had.
- 300
- US CISOs and CSOs
- 500+
- employees per company
- 36
- questions, fielded by Cint
- 24
- pages
- Confident, yet breachedWhat leaders say vs what they found
- Security gets the blameWho owns agents vs who answers for a breach
- Built for loginsShared keys, borrowed credentials, one check per session
- It's already going wrongIncident types, detection, and the industry cuts
- Nobody can keep up by handPolicy volume vs the people writing it
- Trust, within limitsHow far CISOs would let the system decide
- Three times as manyWhere governance is today and in two years
Confident at the door. Blind inside the building.
The same companies. Two very different answers.
Three in four security leaders are very confident they know how many agents are connected to their systems. Seven in ten say security signs off before anyone connects an AI tool.
Yet 49% found an agent running in production this year with access it should never have had. Among the very confident, the incident rate was no lower than anyone else's. Only 34% get their agent count from a tool that discovers agents automatically.
Read the full chapterSource: Agen.co, The State of Autonomous Agent Governance, 300 US CISOs and CSOs, Sept 2026.
Shared keys carry the widest incident gap in the survey.
53% of companies run agents on shared or long-lived API keys, and 40% run them on an employee's own credentials, so the logs show a person doing whatever the agent did.
Most check identity once, at registration (26%) or at session start (45%). Only 18% check before each action, the part of the line where an agent takes thousands of steps a week.
See how identity is checked todaySource: Agen.co, The State of Autonomous Agent Governance, 300 US CISOs and CSOs, Sept 2026.
Three times as many CISOs want governance that runs itself.
We described six ways to govern agents in plain language, from a person deciding every request to a system deciding every action inline, and asked where companies are today and where they want to be in two years.
Over half sit at the bottom two levels today. The share at L3 and above, where the system manages its own policy, rises from 10% to 31%. 44% want to move up at least one level.
Get the six-level modelSource: Agen.co, The State of Autonomous Agent Governance, 300 US CISOs and CSOs, Sept 2026.
83% would let the system decide, inside lines they draw.
We asked how far CISOs would trust an AI system to make agent access decisions in real time, within limits they set, with no human reviewing each one.
23% would trust it fully and 60% would within tight limits. Just 1% would not trust it at all. Companies that had an incident were more than twice as likely to trust it fully (32% vs 14%).
Read what autonomous governance has to be- 23%Would fully trust it
- 60%Within tight limits they set
- 16%Only with a human reviewing
- 1%Would not trust it
Source: Agen.co, The State of Autonomous Agent Governance, 300 US CISOs and CSOs, Sept 2026.
Four lines the data keeps drawing.
Govern the action, not just the login.
Decide at the moment, not at the quarterly review.
Give every agent its own identity.
Shared keys are nobody's identity.
Autonomy inside limits people set.
Never trust without them.
A named person is accountable.
"The agent did it" is not an answer.
The agents are already working without anyone in the loop.
Sooner or later, governance will have to as well. Get the full findings, the industry cuts, and the model for getting there.
Methodology300 CISOs and CSOs at US companies with 500 or more employees answered 36 questions between September 24 and 27, 2026. Cint, an independent research firm, fielded the survey and verified respondent titles. Agen.co wrote the questions and had no access to respondents. Every company in the sample is using, piloting, or evaluating AI agents. All findings are self-reported.