AI agents have moved past answering questions. They now approve transactions, change code, reach into customer data, and take action across business systems, often with no one checking each step. The question for leadership is no longer whether to give agents authority. It is whether they can see and control what agents do with it.
To understand how executives are handling that shift, Salt Security surveyed 500 US and UK C-suite leaders, in research conducted by Censuswide in September 2026. The results are in our new report, Autonomy Before Control: The 2026 C-Suite Report on AI Agent Risk. The short version: autonomy is arriving first, and control is catching up.
Leaders are ready to hand over the keys
92% of executives would allow AI agents to take at least some independent action. That ranges from routine administrative work to decisions with real financial and operational consequences. Very few want agents limited to making recommendations.
That confidence makes sense. Agents promise speed, and no leadership team wants to fall behind competitors that are moving faster. But authority without visibility is where risk builds up.
Most have already seen AI go off script
77% of organizations that use or have explored AI have experienced an AI system or agent acting outside its intended or authorized behavior. For some, the consequences were minor. For others, the result was operational disruption or a security incident.
This is no longer a hypothetical risk to plan for. Most organizations have already experienced it at least once.
What this looks like in practice
In October, Salt Labs research showed how quickly an agent can be turned against the person it works for. A single malicious email could have hijacked the Manus AI agent and given an attacker a path to the user’s connected email, cloud storage and code repositories. The agent’s guardrail detected the malicious instruction, but it could not stop the agent from acting on it. The issue was disclosed responsibly, has been resolved and is no longer exploitable.
The lesson applies far beyond one product. Detection on its own is not control. When agents act at machine speed, organizations need to know what an agent can reach and what it is actually doing, and they need the ability to stop it.
Security practitioners see the same gap from the inside. In Salt’s separate State of Agentic AI and API Security 2H 2026 survey of 383 security practitioners, only 12.5% said they could trace an agent’s full path from prompt to the data it touched, and only 17.5% could automatically stop an agent behaving unexpectedly.
Control gaps are already slowing AI down
Security concerns are not just a risk line item. They are holding back the business. 64% of executives have restricted, significantly delayed or canceled an AI project because of security, governance or regulatory concerns. Every stalled project is delayed productivity and a return on AI investment that has not arrived yet.
For business leaders, this creates a difficult trade-off. Moving too quickly without adequate controls introduces operational and security risks. Moving too slowly risks losing the productivity and competitive advantages that made AI adoption a priority in the first place. The answer is not less autonomy, but greater visibility and control over how that autonomy is exercised.
An inventory is a starting point, not a control
97% of organizations using AI say they maintain an inventory of their AI agents, yet 19% admit it is not kept consistently up to date. Even a current inventory only shows that an agent exists. It does not show what the agent is connected to, what it is allowed to do, or what it did yesterday.
Agents are often deployed by separate teams and inherit access through existing applications, identities, and APIs. A list built once and reviewed occasionally cannot keep pace with that. Governance has to work continuously, following agents as they gain new tools, integrations and permissions.
What else the report covers
The full report goes further into how executives are approaching AI agent risk, including:
- which high-consequence decisions leaders would let agents make without human approval
- who executives believe should be held accountable when an agent causes harm
- where US and UK leaders differ sharply on autonomy, risk and responsibility
- five priorities for adopting AI agents securely
Download Autonomy Before Control: The 2026 C-Suite Report on AI Agent Risk
Frequently asked questions
What is AI agent governance?
AI agent governance is the set of policies and controls that decide what AI agents can access, which actions they can take without human approval, who owns each agent, how their activity is monitored, and how teams respond when an agent acts outside its intended behavior.
Who is accountable when an AI agent causes harm?
There is no settled answer yet. Our survey asked executives exactly this, and the answers vary sharply by role and by country; the full breakdown is in the report. Whatever an organization decides, accountability should be assigned before an agent is deployed, not after an incident.
Is an inventory of AI agents enough to manage the risk?
No. An inventory shows which agents exist. Managing the risk also requires knowing what each agent can reach, limiting its permissions to what the task needs, and monitoring what it actually does at runtime so unexpected behavior can be stopped.
Securing the agentic path
Salt Security, the leader in agentic and API security, helps organizations close the gap between AI agent autonomy and enterprise control. By securing the entire agentic path, from AI agents and MCP servers to the APIs they call, the Salt Agentic Security Platform gives security teams visibility into agent activity, the ability to identify risky behavior, and the controls needed to respond before risks escalate. The result is a stronger foundation for AI adoption without sacrificing security or business momentum.
Request your demo of the Salt Agentic Security Platform today.
Methodology: Censuswide surveyed 500 C-suite professionals at organizations with 50 or more employees, 250 in the US and 250 in the UK, between September 10 and 18, 2026.
