The Wolf Pack Power Brief ·
Agents Without Badges: the AI security gap that will decide 2027 budgets
A OneTrust survey of 1,200 senior leaders shows AI agents getting system access without governance, and what to put in place before 2027 budgets.
- By Paulo Veloso
- 4 min read
Last Monday OneTrust published a survey of 1,200 senior leaders at companies with $100M or more in revenue. One line stopped me.
87% of those organizations encourage employees to use AI agents. Only 47% say they have clear governance, oversight and controls in place to support that.Read that as an operator. A new workforce has been hired, given system access and put to work, and in more than half of companies nobody issued it a badge.
I've spent my career on the revenue side of security, and gaps like this usually go one of two ways. They become next year's breach headline, or next year's budget line.
This one is already turning into both.
Four reports, one pattern
I pulled together the four 2026 studies I trust most on this topic. They were run by different firms on different samples, and they reach the same conclusion.
OneTrust , AI-Ready Governance Report (Sept 2026, 1,200 leaders, 8 countries)
- 86% had at least one AI-related incident in the past year.
- 28% had two or more incidents where AI systems or agents took actions nobody approved.
- Only 27% slowed or paused AI deployment after an incident.
- 33% saw employees turn to unapproved AI because the approved tools didn't arrive fast enough.
Gravitee , State of AI Agent Security (April 2026, 750 US and UK technology leaders)
- An estimated 48% of production AI agents are running unsecured.
- Only 7.2% of organizations have a named individual formally accountable for their agents.
- Confidence in agent visibility rose to 91.8%, while actual monitoring coverage stayed flat at around half.
Okta , AI Agents at Work (May 2026, 292 executives and 492 knowledge workers)
- Only 34% apply the same security controls to agents that they apply to human employees.
- 90% of executives are confident in their visibility into AI tools, while 52% of employees admit to using AI tools without approval.
NeuralTrust , State of AI Agent Security (June 2026, 160+ CISOs)
- 72% are implementing or scaling agents. 29% report comprehensive security controls.
- Only 19% are red-teaming their agents.
The pattern is simple. Confidence keeps going up while control stays flat, and the distance between them is where incidents happen.
Why "unapproved actions" is the number that matters
Most of the AI risk conversation in 2025 was about data going out: someone pasting a contract into a chatbot. That still matters, but agents change the risk. An agent doesn't just read. It acts. It opens tickets, moves money, changes permissions and emails customers.
When 28% of large organizations report repeated incidents of AI taking actions nobody approved, the question is no longer about acceptable use policy. It's about identity and authorization. What is this thing allowed to do, as whom, and who answers for it?
That's a question security teams already know how to answer for humans. We just haven't asked it of agents.
Give every agent a badge
In the sales organizations I've led, nothing reached commit without five answers: access to power, the procurement process, friction, the technical win, and the help needed to close. No answers, no commit. It sounds rigid, but it made the forecast mean something.
Agents need the same gate. Before any agent reaches production, get five answers:
- Who owns it? A named human, not a team or a channel. Today only 7.2% of organizations can say that.
- What identity does it run as? Its own identity, scoped to the task. It shouldn't borrow a person's credentials or an old service account with admin rights.
- What can it do without asking? Write down the actions that need human approval before execution. Anything that moves money, changes access or leaves the company should be on that list.
- How do we see it? Log every tool call and every data access, and send those logs to the same place your security team already watches.
- How do we stop it? A kill switch that someone has actually tested, not one that exists only in a runbook.
None of this needs a new category of product to get started. It needs an owner, a checklist and the discipline to enforce it. The tooling helps you scale it.
Follow the money
This is the part for security vendors, MSSPs and the revenue leaders who sell for them.
OneTrust found that 98% of organizations plan to increase their AI governance technology budgets next financial year, by an average of 25%. And 80% say their teams already spend more time managing AI risk than a year ago, with working hours up 26% on average.
So there's budget, and the people who own it are overloaded. That combination decides who wins.
The same report also contains a warning: a third of companies saw shadow AI grow because approved tools were too slow. Governance that slows the business down doesn't reduce risk. It pushes the risk somewhere you can't see it.
The vendors who capture this budget will sell speed with control.Their pitch will be that they get your agents into production faster, each one with a badge, an owner and an audit trail. Sellers who lead with "we block AI" will get a polite meeting and no deal.
If you sell into this market, expect three questions in every RFP before the end of the year:
- How do you issue and scope identities for non-human workers?
- How do you enforce approval for high-risk agent actions in real time?
- How do you prove, to an auditor, what an agent did and on whose authority?
If your team can't answer those crisply today, that's your enablement plan for Q4.
Let's go build something that sells.
Sources: OneTrust 2026 AI-Ready Governance Report (Sept 14, 2026); Gravitee State of AI Agent Security 2026 (April 2026); Okta AI Agents at Work 2026 (May 27, 2026); NeuralTrust State of AI Agent Security 2026 (June 25, 2026).
Also published on LinkedIn .