Among the 70 security leaders and practitioners surveyed in Tines' Security Hype Check, 91% plan to increase AI autonomy in 2027. Just 9% describe their current AI use as heavily integrated, with autonomy to act on some decisions.
The split is clearer when you look at what teams use AI for. 64% use it for threat detection or anomaly identification. Only 26% use it for containment, such as isolating a host or disabling an account. That makes detection 2.5 times more common than containment. These findings are self-reported and describe the respondents, not the industry as a whole.
One scenario shows why the gap exists. Your security operations center (SOC) detects suspicious activity involving a service account that supports payment processing. An AI agent recommends disabling it. The activity could indicate a compromise. Disabling the account could also interrupt payments.
Should the agent act without approval?
The agent could be right about the threat and still take an action the business has not authorized. Confidence in its assessment isn't enough. Acting requires agreed conditions for intervention, including the disruption your organization will accept. Many teams haven't settled those conditions, which is why detection is often automated and containment often isn't.
The survey points to several barriers, including audit explainability, resourcing and compliance. Three questions sit behind many of them: what the agent is allowed to do, whether anyone can prove what it did and whether the business can recover if it's wrong.
Reason 1: Nobody has defined acceptable disruption
For the payment account, your team needs to know which processes depend on it and who owns the service. A risk score cannot tell you whether disabling the account will pause a background task or prevent customers from completing purchases.
Without that knowledge, a cautious team has good reason to keep a human in the loop. Moving past that caution means security and service owners agreeing on the circumstances that permit immediate containment and those that require approval. That agreement must cover what happens when evidence is incomplete or the designated approver is unavailable. Silence should not become accidental permission.
Those boundaries should be enforced through the agent's permissions and execution controls. A model's confidence cannot determine how much business disruption is acceptable, even when the threat appears urgent.
Reason 2: Teams can't prove what happened, or why
The survey's most common barrier to more AI autonomy was difficulty explaining an AI decision during an audit or review (23%). The payment account shows why.
Suppose the agent disables the account and payments fail. During the review, the agent produces a convincing explanation: the activity resembled credential compromise, so containment was necessary.
That explanation does not establish whether the action complied with policy.
To answer that question, teams need records captured during execution showing the evidence available to the agent, including its source and freshness. Reviewers should be able to identify the policy version applied and any approval obtained. They also need to distinguish what the agent requested from what the system did. An accepted API request does not prove containment worked. The team must verify the intended access was blocked, including whether existing sessions remain active.
A generated summary can help reviewers understand those records. It cannot fill gaps in the evidence.
Someone also needs to own exceptions: investigating them and suspending the agent's authority when necessary. If the agent was right about the compromise but acted outside its permissions, better detection won't fix the control failure.
Reason 3: Recovery is harder than the undo button
Re-enabling the account may take seconds. Restoring the payment process may take much longer.
Failed transactions may need reconciliation. Retries could create duplicates. Customers whose purchases failed may already have gone elsewhere. A reversible technical action can still cause lasting business harm.
Teams that have thought this through are right to be wary. Before granting autonomous authority, test these scenarios with the payment service:
- A genuine compromise
- A false alarm
- Stale account information
- An unavailable approver
- A containment request that only partly succeeds
Verify that the agent stops or escalates when the policy's conditions are missing. Then rehearse recovery with the service owner. Confirm who can halt further agent actions and how the team will restore access without reopening the compromise. Measure recovery until payments work correctly again, rather than stopping the clock when the account is enabled.
Use those results to define when immediate containment justifies the disruption, when a narrower action is appropriate and when approval is required.
What it takes to extend autonomy
Hesitation about autonomous containment isn't a failure of nerve. Budget and regulatory constraints also rank high among the barriers (22% each), but in many cases the hesitation reflects questions that haven't been answered yet. Before letting an agent act, a CISO should be able to say:
- What the account supports and who owns it
- Which policy permits intervention
- How the team will verify the action worked
- How long recovery takes and who leads it
The payment account might need immediate disabling. It might require human approval. The scenario alone cannot settle that choice. AI confidence can inform the threat assessment; permission must come from the agreed response policy.
Read the full Security Hype Check report to see what 70 security leaders and practitioners said about AI autonomy.