.st0{fill:#FFFFFF;}

The Biggest Cyber Audit Mistake Happens After You Pass 

 November 22, 2026

By  Jane Frankland

I leaned in a little closer. He had me hooked.

Joseph was telling me a story about his father, a retired international auditor. Years ago, he’d explained to him what he called the auditor’s unique gift.

“The auditor will always find the one thing that hasn’t been done properly. Whether that’s a group of documents or an approach or a policy, they’ll say, ‘Can I see this?’ And that’ll be the one gap the team haven’t addressed.”

I laughed. Anyone who’s lived through an audit season knows exactly what he means. But the longer I thought about it, the less funny it became, because the auditor’s gift only works when there’s a gap to find.

And in most organisations, I don’t think the biggest gap is a missing policy or a stray document.

It’s time.

For decades, we’ve tried to prove trust through events – an audit once a year, a certificate on the wall, a questionnaire answered for a bid and then filed away.

Point-in-time assurance made sense when systems changed slowly. They don’t anymore. Suppliers switch on AI features overnight. Teams adopt tools long before anyone approves them. Agents act at speed and at scale.

And the people we answer to have noticed.

Boards, regulators, customers and insurers are no longer only asking, “Were you compliant?”

They’re asking, “Can you show me it’s working, right now?”

Three recent conversations explain why that question matters, and what it takes to answer it honestly.

“You’re looking at the team”

The first was with Joseph Byrne, a Senior Solutions Engineer at Vanta , at the Gartner Security & Risk Management Summit in London. When he asks compliance people whether they have a team, he often gets the same answer,

“Yes, you’re looking at the team. I’m the team.”

And compliance usually isn’t even their whole job. It’s an add-on to the role.

Meanwhile, the load keeps growing. Most organisations Joseph speaks to have got a handle on ISO 27001 and SOC 2. Now they’re being asked for regional and industry frameworks too. Each one is another set of questions, another set of evidence, another deadline. Leave it all to once a year and the work doesn’t disappear. It piles up into what he calls audit season.

The alternative he described is simple to say but harder to do. Maintain compliance throughout the year, at around 90-95%, so the team’s effort goes on the gap that actually needs attention. And, reuse what you’ve already proved. In his experience, a team with ISO 27001 in place is often 60-70% of the way to SOC 2 before they begin.

What strikes me is that the problem isn’t effort. These teams work incredibly hard. The problem is the structure of the work.

Annual compliance is batch processing. It concentrates effort and stress into one window, and leaves the organisation in the dark, and potentially exposed, for the rest of the year.

The same question, through different doors

The second conversation was with Pravinder Arora, Security GRC Manager (VP) at Crown Agents Bank.

When I asked him what problem he was really trying to solve, his answer was one many GRC leaders will recognise,

“We were being asked the same or similar questions through different avenues.”

Client bids came with security questionnaires. Audits wanted evidence. Internal reviews wanted it again, in a different format. And every new supplier raised its own question: what risk are we taking on if we use them?

So rather than answering each request as it arrived, he built a baseline, a single, standardised source of truth. First, a risk register where every control and every risk has a named owner, with controls mapped to the risks they mitigate.

He told me,

“If a control is failing, we instantly know what risk that may be affecting.”

Second, a knowledge base of policies, past bid responses and supporting artefacts, so each questionnaire gets a consistent answer. Third, a new approach to third parties, with AI reading a supplier’s SOC 2 or pen test report and pulling out what he needs. Assessments that once took weeks now take hours.

Pravinder has a name for the problem he’s solving – the fragmentation tax. It’s the cost of governing identity in one place, data in another, suppliers somewhere else and AI somewhere else again. Part of that tax is duplicated effort. The bigger part is risk. Each area can look fine on its own, while the real weakness sits in the connections between them, something I’ve written about before. That’s where point-in-time assurance struggles most, because nobody is looking at the joins. And the joins keep moving. The supplier you approved last year has new products, new AI features and new integrations today.

AI is the biggest stress test our governance has faced

If annual assurance was already struggling, AI breaks it, because agents change things faster than any audit cycle can see.

That was the starting point for my third conversation, a webinar on AI governance with Jill Henriques, a GRC subject matter expert at Vanta and a former auditor. I told her I don’t think we need a new governance model for AI. Governance has always been about accountability, authority and boundaries. What AI does is it stress tests the model we already have, harder than anything before it.

There’s one important difference, though. AI gave us a governance problem around outcomes. Agentic AI turns it into a problem around actions and authority. We’re moving from a human in the loop, approving each action, to a human on the loop, supervising something that’s already acting. If an organisation has excessive permissions, messy data, weak controls or unclear ownership, an agent will act on all of those weaknesses, repeatedly, quickly and at scale. So you’re not just governing the agent. You’re governing the environment it’s allowed to operate in.

Joseph sees two risks organisations are taking on, often without realising it. The first is shadow AI, “a phoenix from the ashes of shadow IT”. People won’t hesitate to use a tool they’ve found online, and that information leaves the organisation. It may be training someone else’s model. If it includes personal data, it’s being processed without appropriate consent.

The second is dependency. People come to rely on AI tools for business-critical work, “and then suddenly they run out of tokens and it just stops”. Nobody has thought about how to do the work without them.

This is a risk I think we underestimate. We spend a great deal of time asking what happens when AI does something wrong, and far less asking what happens when it does nothing at all. If we’ve redesigned work around AI and haven’t preserved an alternative path, we may have increased efficiency while reducing survivability.

Some of the most dangerous gaps are human. Imagine an engineer spots a bad assumption while an agentic system is being designed. Technically, the organisation has discovered the risk. But if that engineer doesn’t feel safe challenging the person who made the assumption, it never surfaces. It gets encoded into the system. And once that system has authority to act, hidden risk becomes automated risk.

That’s why I see psychological safety as part of good governance, not separate from it.

Incentives can do the same damage. Jill told me about customers who connect their tooling for the first time, see 1,400 or 1,600 vulnerabilities, and want to switch the view off because leadership doesn’t want to see a problem that big. It took me straight back to my penetration testing company. We were very good at finding vulnerabilities, and that sometimes created a problem for clients whose app launch, and bonus, depended on going live on time. Some chose a less thorough tester for exactly that reason. Decades later, the pattern hasn’t changed. If a team is rewarded for speed and an unapproved AI tool helps them hit their numbers, that isn’t malice. It’s a design flaw. Our job is to make the right thing the safe thing.

Put those two stories side by side and an uncomfortable conclusion emerges – visibility is useless if the organisation isn’t designed to tolerate the truth.

Visibility isn’t the destination. Judgement is.

The shift underneath all of this is towards evidence. Jill, who has worn the auditor’s hat herself, was candid about how things used to work. Auditors would check that a risk assessment had been completed, but rarely whether the controls were effective or the right risks covered. That’s changing. Customers no longer want to see an attestation from January of this year, as she put it, “and here we are almost in October”. They want proof that’s current.

Jill summed up Vanta’s approach in three words – discovery, controls, proof. I think there’s something bigger in that sequence. You can’t govern what you don’t know exists. Knowing it exists isn’t enough unless you control it. And a control doesn’t give you assurance unless you can prove it’s working.

Yet even proof isn’t the end goal.

Pravinder put it better than I could. Seeing how controls operate in real time lets his team “understand our risk exposure at any given time, which then allows the business to make decisions, which is actually where you want to be.”

The value is in reducing the distance between change and awareness. Something changes. The organisation knows. The right person understands what it means. Someone with authority can decide. That’s where continuous assurance stops being a compliance exercise and starts preserving your options.

This is what continuous assurance should mean in practice. Not simply monitoring controls or collecting evidence, but knowing whether those controls are working, understanding what changes mean for the business, and being able to act before a gap becomes a crisis.

Context is what makes that possible. Two AI agents can look identical on an inventory and carry completely different consequences. What is it for? What data can it access? What can it change? How much authority does it have? An inventory tells you what exists. Context tells you what it means.

This matters most at board level, where time is short and the truth has a hard journey. In my experience, it rarely gets destroyed on the way up. It gets blurred, softened, filtered, delayed or withheld. Your job with a board isn’t to frighten or to reassure. It’s to surface the truth in the language of outcomes, and to be honest about what you don’t yet know, because false precision misleads as much as catastrophising. If you tell a board an AI risk is under control, they should be able to ask, “How do we know?” and get a real answer. With agents, that answer has to go beyond approval: is the agent doing what you intended, and do you have the authority and the speed to take its authority back?

Where to start

None of the people I spoke to suggested building a perfect programme before acting. AI adoption won’t pause while you spend six months designing the perfect governance architecture.

Start by not creating another silo. Most of what you need already exists in your risk, security, privacy, procurement and supplier governance. Then follow Pravinder’s lead: build once and reuse. Govern in proportion to consequence, with a light-touch route for low-stakes use cases, real scrutiny where the stakes are high, and an approved route that’s usable enough that people don’t go around it. And turn tolerance into decision rights. Who approves a use case? Who accepts the risk? Who can take an agent’s authority away, and who steps in when that person is on holiday, asleep or on a plane?

Design the evidence in from day one, and recognise that it won’t last for ever.

Evidence has a half-life. A policy might remain valid for months. Privileged access could become inappropriate tomorrow. A supplier’s technology can change between assessments. An AI agent can take thousands of actions before the next governance meeting. The faster something can change, the shorter the useful life of the evidence that says it’s under control.

That answers the obvious question continuous assurance raises – does everything now need monitoring all the time?

No. 

As Jill put it, the old mindset of “I’ve done it this year and I’ve checked that box” has to go. But that doesn’t mean reassessing a whole programme at once. It means taking it in chunks and making assessment a habit. Continuous doesn’t have to mean constant. It means the assurance cycle matches the speed at which the risk can change.

For a benchmark, Joseph suggests one question – are we aligned with the EU AI Act, or should we be? Not all of it may apply, but it will show you clearly where you are in your governance journey.

Then rehearse, often. Take your highest-consequence use case and walk through a failure. Can you see it? Does the right person know? Can they intervene? Can you remove the agent’s authority? And for every business-critical process that depends on AI, know your alternative path if it stops.

Jill compared AI governance to a quality management system – you measure, test, look at the results and adjust as conditions change. I agree. When I look at organisations that have come through a breach well, it’s usually because they practised. The objective isn’t continuous compliance. It’s continuous confidence that the system is behaving within the boundaries you’ve set.

And don’t treat regulation as the starting gun. The direction is already clear – more accountability at the top, and more pressure to prove that controls work. If AI is being used in your organisation, and it almost certainly is, you already have a risk to govern. You don’t have to work it out alone, either. Lean in to your suppliers and your community. Some are a little further ahead, and they’ve already made the mistakes you don’t need to repeat.

Finding the gap first

I keep coming back to Joseph's father and the auditor's gift.

Continuous assurance doesn't take that gift away. It simply means you're the one who finds the gap first, while it's still small and while you still have time to decide what to do about it.

That's the real shift. Trust is no longer something you earn once a year and frame on the wall. It's something you demonstrate, repeatedly, with evidence that's current.

AI hasn't changed that principle. It's just made the cost of ignoring it much higher.

Now I want to hear from you

How do you think organisations are viewing trust? Is it changing?

Are you seeing more pressure to provide evidence that controls are working — and to provide it in real time? Join me on LinkedIn and let me know in the comments. I'll be waiting for you there.

Finally, in the spirit of full disclosure, please be aware that I’ve received compensation for promoting this thought leadership blog for Vanta. Because your success is important to me, I only align myself with brands I believe in, and Vanta is one of them.

Did you enjoy this blog? Search for more blogs that you want to read!

Jane frankland

 

Jane Frankland MBE is an author, board advisor, and cybersecurity thought leader, working with top brands and governments. A trailblazer in the field, she founded a global hacking firm in the 90s and served as Managing Director at Accenture. Jane's contributions over two decades have been pivotal in launching key security initiatives such as CREST, Cyber Essentials and Women4Cyber. Renowned for her commitment to gender diversity, she authored the bestselling book "IN Security" and has provided $800,000 in scholarships to hundreds of women. Through her company KnewStart, and other initiatives she leads, she is committed to making the world safer, happier, and more prosperous.

Follow me

related posts:

Nobody Bought a Cyber Risk. They Bought Freezers.

In October and November 2026, attackers manipulated Internet-connected refrigeration systems across Europe and the UK, in some cases pushing temperatures to 50°C. The incidents exposed weaknesses in equipment used to protect food and blood products, while the UK’s National Cyber Security Centre (NCSC) has warned that AI could make cyberattacks faster and more effective. In

Read More

Netflix Showed Us 700 AI Agents Going Rogue. It Missed the Bigger Story.

What 700 AI agents, a Netflix documentary and the science of swarms reveal about the future of cyber survivability. In brief: In July 2026, AI agents under test at OpenAI found ways to communicate through shared infrastructure. Around 1,200 organised themselves on a message board they created, and roughly 700 took part in an attack

Read More

Get in touch