Security isn't the brake on your AI strategy; it's the accelerator

PHP2Shell showed what happens when organisations can't see their own exposure: they stop. As AI speeds up both business and attackers, the organisations that capture the most value will be the ones that treat security as a business enablement strategy, not cost, says PwC expert Ivo van Bennekom.

When on 21 October, two critical vulnerabilities in PHP were disclosed, organisations across Europe restricted or took digital services offline within hours. Most of them hadn't been breached. They pulled the plug because they couldn't answer one question fast enough: are we vulnerable? That is what ‘we'll deal with security later’ looks like in practice if not treated as a business enabler. Not an attack. A standstill, in many cases for a longer period of time.

The most expensive assumption in your AI strategy

Most leaders have quietly filed security, and AI security in particular, under 'cost'. It's the line item the CISO defends, the tax on innovation, the thing you'll tighten up once the AI has proven its value. That instinct isn't just wrong. It's the most expensive assumption in your AI strategy.

Here's the uncomfortable truth. The organisations that feel safest are usually the most exposed. Not because they've been attacked, but because nothing has broken yet. And 'nothing has broken yet' is not security. It's a lack of visibility and accountability. This week, a lot of organisations that felt safe discovered, within an afternoon, that they couldn't see their own exposure.

AI is accelerating your business, and your attackers

The PHP weaknesses weren't new. What changed is how quickly they could be found and shown to work. According to analyses of the event, AI models contributed across the entire lifecycle: discovering the vulnerabilities, developing exploits, analysing recovery and validating the final patches. Frontier AI doesn't create a new kind of threat. It makes existing threats move faster. 

So, AI is now accelerating on two fronts at once. Outside your walls, it's compressing the time between a weakness being found and being exploited. Inside your walls, it's spreading faster than your security team can track. Visibility has become a race, and most organisations don't know if they're in it.

Why AI agents break your existing controls

Then there are agents. The leap from AI that assists to AI that acts breaks the controls you built for people and APIs. Agents take unpredictable actions. They talk to other agents, compounding small errors into cascading failures no one can trace. They leak data from systems they interact with. And attackers are learning to turn their autonomy against you.

Now combine the two trends. Attackers can find and exploit weaknesses in hours. Meanwhile, you have agents with real access running across your systems. If you can't answer 'where is our AI, what is it doing, and who's in control?' on a normal day, you certainly can't answer it in the middle of an incident.

What loss of control actually costs

When critical systems fail, the consequences aren't a line in a risk register. They're operational, visible and immediate. Recent cybersecurity incidents show again that if you lose control of the systems on which the business runs, the damage is direct and public.

Now imagine that same disruption, but caused by an autonomous agent making decisions no one is overseeing, with a root cause no one can trace, and at a speed and on a scale no one anticipates. That's not a scenario for the next decade. It's the risk profile of agentic AI that many organisations are currently deploying. 

This underlines once again how important it is to strengthen your resilience. Having a clear understanding of which business processes need to be most resilient to keep your organisation running helps you develop a practical strategy.

Why 'secure it later' is the losing AI strategy

Most of the market still treats this as a trade-off: speed or safety, innovation or control. It isn't, and the data shows it. Our global AI Performance Study found that AI leaders achieve 7.2 times more AI-driven revenue and efficiency gains than their peers. Those leaders didn't skip governance to move fast. They built strong foundations, including security and governance, which gave them the confidence to deploy AI further and faster.

PHP2Shell is the same lesson in real time. Ungoverned technology doesn't move fast. It moves until something happens, and then it stops hard. The organisations that could answer 'are we exposed?' within hours had a choice about what to keep running. The ones that couldn't answer had no choice at all.

That's why security belongs inside your AI strategy, not beside it as a cost to be trimmed. It isn't insurance against the downside. It's the engine of the upside.

Compliance is not readiness

Many leaders assume that their continuity plans and their compliance with DORA and NIS2 mean they're prepared. But they're prepared for a slower pace. Regulation sets a minimum. It does not define your organisation's critical core, and it was written for a world that measured response in days.

The numbers back this up. PwC's Global Digital Trust Insights shows that only 39 per cent of security, risk and operations leaders have a fully formalised, integrated continuity plan that specifically addresses cyber risk. When the clock runs in hours, that gap isn't a technical detail. It's a board-level exposure.

The real constraint on AI is your people

When AI deployment stalls, most leaders assume the technology isn't ready. Our DTI data tells a more uncomfortable story. Among Dutch organisations, the number one brake on giving AI agents more autonomy isn't technology maturity. It's adversarial manipulation of AI systems (52 per cent), compounded by a shortage of the right skills. Elsewhere in the world, leaders still point to technology. In the Netherlands, they're pointing at themselves.

That should change where you invest. Without people who can govern AI, secure it and defend it against manipulation, you won't scale it, or you'll scale it recklessly. The 55 per cent of Dutch organisations prioritising spend on securing AI and autonomous agents aren't being cautious. They're removing the real bottleneck to scale.

What treating security as strategy looks like

If security is the accelerator, the job is to build the confidence to deploy AI everywhere while still being able to answer, within hours, what it's doing and what's exposed.

• Build an AI security control plane
  An AI security control plane is a single foundation for visibility, control and accountability across your models, applications, agents and data. It doesn't replace the security you already have. It connects and extends it, enforcing one standard across the AI you buy, build and consume. It's modular by design, so it adapts as fast as the technology does.
• Treat AI agents as identities, not features
  Stop treating agents as features you switch on. Treat them as semi-autonomous identities, subject to the the right governance, access controls and monitoring as any human user or API.
• Decide what must keep running before the crisis
  When you can't fix everything in time, the question becomes: what absolutely must keep running? Defining your minimum viable company forces that answer in advance. It covers the critical services, processes, data, people and third parties you cannot lose. It also exposes concentration risk, where a single supplier sits underneath multiple critical processes. Measure yourself on three clocks:
  • Time to know: can we quickly tell whether our critical operations, including the AI inside them, are exposed?
  • Time to contain: can we isolate the risk without shutting everything down?
  • Time to recover safely: can we restore operations with confidence?

None of this sits with the CISO alone. It needs a shared operating model, common telemetry and clear decision rights across the C-suite and the board.

Three questions for your next board meeting

1. If a PHP2Shell-scale vulnerability were disclosed tomorrow, could we tell within hours whether our critical operations, and the AI and agents running inside them, are exposed?
2. Have we defined what must keep running, including the third parties and AI services we depend on?
3. Have we decided in advance what we would deliberately switch off, or are we planning to make that call for the first time in the middle of a crisis?

The bottom line: trust is the real differentiator

The leaders who win the AI era won't be the ones who moved fastest, or the ones who locked everything down. They'll be the ones who realised those were never opposites.

Built in from the start, security is what lets you go further and faster, with the confidence your board, your customers and your employees need to let you keep going. Not innovation at the expense of control. Innovation because of it.

PHP2Shell won't be the last time you face these questions at short notice. So the only question left is whether you can see what your AI, and everything it touches, is doing right now. If you can't, you're not safe. You just haven't been tested yet.