Reviewed by: Mansoor Ali, Technical Editor, PenPonder | Last Updated: July 2026
A new survey of 1,001 IT and security leaders found something that should worry any business using AI right now. 78% of organisations have already had an AI-related security incident or found an AI-related vulnerability. That number is not about AI writing bad code. It is about something more basic. Most companies gave their AI systems real access to real data, then never gave them a real identity.
This guide covers what the data actually shows. It covers why the usual fixes miss the point. And it covers what to check first if you are using AI tools in your business right now.
The Survey Behind This
DigiCert published this research in its AI Trust Outlook report. Propeller Insights ran the underlying survey in May 2026, on DigiCert’s behalf. It covered 1,001 IT and cybersecurity decision-makers across the United States, United Kingdom, and Australia.
The headline number is 78%. That is the share of organisations reporting an AI-related incident or vulnerability. Broken down further, 27.7% had exactly one incident. 21.9% had multiple incidents. 28.4% found a vulnerability but had not yet suffered an incident from it.
A separate report from identity security company CyberArk adds useful scale to this picture. Machine identities, which now includes AI agents, already outnumber human identities by more than 80 to 1 inside a typical organisation. Yet 88% of organisations still define “privileged” access as something only humans hold. That mismatch, treating a workforce that already outnumbers humans 80 to 1 as an afterthought, is a large part of why incidents keep happening.
The Part That Actually Matters
Here is the detail most coverage of this survey skips. The incidents were tied to unauthorised or misconfigured AI agents. They were not caused by flaws in AI-generated code. This is an important distinction. Most security advice about AI focuses on checking the code an AI writes. This data points somewhere else entirely.
| Finding | What DigiCert Found |
|---|---|
| Organisations with an AI incident or vulnerability | 78% |
| New AI systems deployed in the last 6 months | 75% deployed 4 or more |
| AI agents given a unique digital identity | Only 40% |
| Cannot trace an AI decision back to its source | 47% |
| Discuss AI governance at board level | 90% |
| Actually have budget and a formal programme for it | 50% |
Look at the gap between the last two rows. 90% of organisations are talking about AI governance in the boardroom. Only half of them have actually funded it. That gap is where the 78% incident rate comes from.
Why “Give It an Identity” Is the Real Fix
An AI agent that reads your email, updates your CRM, or approves a workflow is doing something a human employee used to do. A human employee has a login. Permissions tied to their role. A record of what they accessed and when. Most AI agents right now have none of that. They often run with broad, shared access, under a generic service account nobody watches closely.
Brian Trzupek, Senior Vice President at DigiCert, put it directly in the report. The real question is no longer whether a business should use AI. It is whether that business can actually explain, govern, and trust the AI it has already deployed.
Only 40% of organisations have given their AI agents a real, individual digital identity, separate from a shared account. Without that, you cannot answer basic security questions. Which agent touched this data. Under whose authority. What permissions did it actually have. When an incident happens, and 78% of companies say one already has, those are exactly the questions you need to answer fast.
Want a more technical framework for thinking through this? OWASP’s Top 10 for Large Language Models is the closest thing the security industry has to a shared standard right now. It is a free, vendor-neutral resource. It covers exactly the risks this article describes, including agents given more permission than they need, and systems that can be tricked into misusing legitimate access.
The Traceability Problem
47% of organisations cannot fully trace an AI system’s decision back to the model and data that produced it. This is often called the black box problem, and it is not just an abstract concern.
If an AI agent takes an action you did not expect, you need to know why. Maybe it denies a transaction it should have approved. Maybe it exposes data it should not have touched. Without traceability, you are stuck guessing. That makes it much harder to explain what happened to a regulator or a customer.
What to Actually Check First
If your business already uses AI tools, four checks are worth doing this week, not next quarter.
Count your AI agents like you would count employees. List every AI tool with access to your systems, your data, or your customer information. Most businesses have more of these than they realise, since teams often adopt AI tools individually without a central record.
Check who actually has a unique identity, and who is sharing one. If several AI agents are running under one generic service account, you cannot tell them apart when something goes wrong. This is the single most common gap in the DigiCert data.
Ask if you can trace a decision back to its source. Pick one AI system your business relies on. Try to answer, in writing, what data it used and why it produced a specific output. If you cannot answer that clearly, you have found your traceability gap.
Match your budget to your board conversation. If AI governance is being discussed at the leadership level but has no dedicated budget or owner, that gap is exactly where the DigiCert data says incidents come from.
For the broader picture of how AI is reshaping cybersecurity work more generally, see our AI in Cybersecurity guide. For the foundational compliance requirements that AI governance now sits alongside, see our Cybersecurity Compliance guide.
Frequently Asked Questions
What percentage of businesses have had an AI security incident?
78%, according to DigiCert’s AI Trust Outlook report, based on a May 2026 survey of 1,001 IT and cybersecurity decision-makers across the US, UK, and Australia. This breaks down to 27.7% with one incident, 21.9% with multiple incidents, and 28.4% who found a vulnerability without a confirmed incident yet.
Are AI security incidents mostly caused by bad AI-generated code?
No, according to this data. The incidents were tied to unauthorised or misconfigured AI agents, not flaws in code that AI wrote. This points toward an identity and access management problem rather than a code quality problem.
What does giving an AI agent a digital identity actually mean?
It means treating an AI agent the way you would treat an employee. Give it a unique login. Give it specific permissions tied to what it actually needs to do. Keep a traceable record of its actions. Right now only 40% of organisations do this. Most AI agents instead run under shared or generic access, which makes it far harder to see what any specific agent did if something goes wrong.
Why can’t some companies trace their AI’s decisions?
47% of organisations surveyed said they cannot fully trace an AI system’s output back to the specific model and source data that produced it. This happens when AI tools are adopted quickly without building in logging, documentation, or oversight from the start. It becomes a real problem when an AI system does something unexpected and a business needs to explain why.
Is AI governance actually a budget problem or an awareness problem?
Both, but the data suggests awareness is well ahead of funding. 90% of organisations say they discuss AI governance at board or executive level. Only 50% have a dedicated budget and a formal programme behind that discussion. That gap between talking about the problem and funding a fix for it is where DigiCert’s researchers link the 78% incident rate.
Data and figures sourced from DigiCert’s AI Trust Outlook report, based on a survey conducted by Propeller Insights in May 2026, independent reporting and analysis of the findings, and CyberArk’s 2025 Identity Security Landscape report. PenPonder does not have a commercial relationship with DigiCert or CyberArk. Survey figures represent self-reported responses from IT and cybersecurity decision-makers and may not reflect every organisation’s experience.
