The instinct to hide, not fix
A new survey from Workato, conducted with Censuswide among 1,001 IT decision makers, puts a hard number on an instinct that has been circulating informally in CIO conversations all year: when an AI agent causes a problem, a meaningful share of leaders would rather manage the story than manage the incident. Roughly a third of respondents said they would try to handle an agent driven security incident quietly, without formal disclosure, if they thought they could get away with it.
That is a striking admission for a survey conducted on the record, which suggests the real number in practice could be higher. It also reframes the governance conversation CIOs have been having mostly in terms of prevention. Prevention matters, but this data says enterprises also need a clear, pre agreed disclosure policy for when an agent causes harm, because the default instinct under pressure is apparently to suppress rather than report.
Approval is the exception, not the rule
The deeper problem the survey surfaces is upstream of disclosure. Nearly two thirds of respondents said most AI agents running inside their organization lack formal IT approval. That means the typical enterprise in this survey has more unapproved agents operating than approved ones, a sprawl pattern that mirrors shadow IT in its earlier, unmanaged years, except these systems can act on their own rather than simply store data somewhere IT cannot see.
Forty two percent of respondents said they specifically want to reclaim visibility into what agents are doing across their workflows, which is a telling way to phrase it. These are not leaders asking for a new capability, they are asking to get back something they already lost. That framing matters for how CIOs pitch governance investment internally: this is remediation of an existing gap, not a speculative new control layer.
Executives are more worried about agents than attackers
Perhaps the most striking figure in the survey is that six in ten respondents say they worry more about AI agents operating inside their own environment than about traditional cyberattacks coming from outside it. For an industry that has spent two decades building its risk vocabulary around external threats, perimeter defense, intrusion detection, and attacker attribution, that is a genuine reversal, and it reflects a real shift in where the unpredictable variable now sits. The thing most likely to cause an incident is no longer assumed to be an adversary; it is a system the organization deployed on purpose.
Almost two thirds of executives go further and expect an AI agent in their own environment to cause a public security or compliance incident within the next eighteen months. That makes the finding a near term expectation held by the majority of the people responsible for preventing it, well past the stage of a hypothetical risk being modeled for a board slide. Few enterprise risk categories get framed with that kind of certainty before the incident has actually happened, and fewer still get surveyed this honestly by the people who would be blamed for it.
What Adam Seligman says is actually going on
Adam Seligman, CTO of Workato, frames the results as evidence that agent governance has simply fallen behind agent deployment at almost every enterprise surveyed. Agents got adopted fast because the productivity case was obvious and the barrier to trying one was low, often lower than procuring a traditional software tool, sometimes requiring nothing more than a credit card and an afternoon. Governance, which requires cross functional buy in from security, legal, and compliance, moves on a slower clock than a team installing an agent to automate one workflow, and that mismatch in speed is the whole story the survey numbers are describing.
The result is the gap this survey quantifies: widespread deployment, limited approval, limited visibility, and executives who already expect the gap to produce a real incident before it gets closed. Seligman's point is less about blame and more about sequencing. Enterprises adopted the easy part first and left the hard part, governance, for later, and later has arrived faster than most planned for. Closing that gap retroactively is harder than building it in from the start, which is the uncomfortable position most of the enterprises in this survey now find themselves in.
How this compares with other 2026 governance data
The Workato findings do not stand alone. The survey cites a New Relic report finding that one in four AI agents operate without oversight in enterprise environments, an EY Technology Pulse Poll finding more than half of department level AI initiatives lacked formal approval, and an UpGuard report finding more than eight in ten employees use unapproved AI tools at work. Taken together, these numbers describe the same underlying condition from different angles: adoption has outrun governance across nearly every measure anyone has bothered to survey.
That convergence across independent surveys is itself useful information for a CIO building an internal business case. A single data point can be dismissed as one vendor's framing of the problem. Four independent studies landing on the same conclusion within the same year is much harder to wave away in a budget meeting, and it gives governance investment the kind of external validation that usually only follows an actual incident.
The policy CIOs should write before the incident, not after
The practical move this data points to is not just tighter agent approval, though that matters too. It is a written, pre agreed incident disclosure policy specifically for AI agents, created before an incident occurs rather than improvised under pressure when one does. The survey's most uncomfortable finding is not that agents might fail, every system eventually does. It is that a third of the people responsible for responding have already told researchers their default instinct is concealment.
A disclosure policy decided in advance, with legal and communications already looped in, changes that default without requiring anyone to make a brave decision in the moment. Pair that with closing the approval gap the survey identifies, and a CIO has turned three uncomfortable numbers into a concrete two part project: get agents under formal approval, and agree now on what happens when, not if, one of them causes a problem anyway.
