Fighting AI threats: How security teams can shift from firefighting to prevention

The current approach to AI in defensive security is becoming an unwinnable race. Attackers use AI to find and exploit weaknesses faster, so defenders use AI to find and fix them faster – at the cost of the long-term planning that would change the odds.

It's an understandable response, but it’s not a sustainable long-term solution, and security teams must re-think their strategy. This starts with understanding what AI has enabled attackers to do and where it falls short in defence. 

In this article, I want to share practical steps for security teams to shift their strategy from firefighting to prevention.

The current flow of AI

AI has lowered the barrier to entry for hostile individuals and groups, enabling those without significant skill to disrupt and potentially destroy businesses.

Research by Contrast Security found that the average application is hit by 42 viable attacks per month, confirmed to reach real vulnerable code. The other factor that makes this concerning is that fixing the most severe vulnerabilities takes an average of 92 days. Taken together, these figures show that defenders are significantly slower at patching vulnerabilities than threat actors are at attacking them. 

Similarly, in areas like Operational Technology (OT), there is an industry-wide expectation that AI-assisted exploitation will rise significantly, according to a piece in CSO. The ability to build working exploits means threat actors can conduct a range of harmful actions, from disrupting operations to causing significant reputational damage.

Thankfully, as the article points out, AI is not currently making zero-day discovery ‘zero-effort’ for low-skilled threat actors, and its use can still be very time consuming and require a lot of manual input to execute attacks. 

With that said, less complicated exploits, particularly publicly available ones, are easier for threat actors to find, modify and use against your organisation. Over time, it’s reasonable to assume that AI capabilities will continue to improve, making them more dangerous in the hands of threat actors. 

If we don’t plan ahead with that in mind, the common vulnerabilities we leave unresolved could be the start of a very long headache.

The impact is also still at times unpredictable. As part of an experiment covered in the article, an AI agent ended up making a device permanently unusable after compromising it. In the world of OT—that controls physical processes—an unrecoverable crash becomes a significant business continuity risk. This makes it even more critical to ensure AI cannot gain access to fragile systems where the results could be significant disruption to operations and physical harm to people.

How does this affect security teams here and now?

While AI’s capabilities increase and more stories hit the media, security teams face mounting pressure to adapt and fix newly identified vulnerabilities at increasing rates. This expanding workload can sometimes be partially offset by introducing AI or automated tools, but even in those cases, the workload doesn’t just disappear. 

AI is good at detecting problems, but using it to judge the severity of those discovered vulnerabilities is where it becomes unreliable, because ultimately you still need human judgement and context. So, rather than being reduced, the workload has shifted and a human security tester is still required to triage the growing queue.

This trend of firefighting is not only unsustainable, but also hinders a security team’s ability to think and plan long-term so they are in a stronger position to handle the challenges posed by AI-driven attacks. Equally, haste in fixing vulnerabilities can introduce problems of its own, such as system instability or inaccessibility.

Starting from the top

Right now, the main train of thought is that the future will become ‘attacker AI vs defender AI’. The odds will not be in favour of the defenders in this battle. 

Defensive AI tools, like AI vulnerability scanners, have been shown to disagree with each other about what the real problems are and lack business context for what vulnerabilities should be prioritised, why and when. As explained by David Linder CISO at Contrast Security, these AI tools, “disagree with themselves from one run to the next”.

This puts AI tools built for defence at a disadvantage compared to AI tools used for attack. Put simply, defensive tools need to work in harmony, from one beat to the next. AI tools in defence influence the problems the security team focuses on and when they point at false positives, this works in favour of the attacker rather than the defender.

In the face of an opponent we have not prepared for, this battle of attrition may buy us time, but on its own, it is not a solution. In the short to medium term, this makes sense on paper, but it misses one crucial element...

Strategy!

Fighting on the frontlines, AI vs AI is not a sustainable strategy. At best it is a short-term battle plan.

For us to genuinely be prepared to defend against the increasing waves of AI-enabled threat actors, we need to start by thinking ahead.

Start with design

How can we design a process, system or physical location that enables us to work as we need whilst minimising the avenues that could be used to cause our business harm? 

Let’s say you're seeking to protect your internal environment. You can start by using network segmentation to limit what a compromised system can connect to, disabling services nobody uses, restricting access to administrative services and constraining what an application is even capable of doing at runtime. The CSO article makes the same recommendation for OT environments specifically. 

Importantly, every avenue removed at design stage is one you never have to patch under pressure at 2am.

Policies and controls

Similarly, how can we incorporate policies and controls that help to prevent or lessen actual vulnerabilities from being present (and exploitable) in the first place? 

• Clear and concise development requirements that limit ambiguity.
• Designing in alignment with key security principles such as ‘least privilege’.
• Security code reviews before code is used in a production environment.

All of these measures reduce the rate of new vulnerabilities and costs are reduced downstream as a result.

Shifting from what's probably to what's possible

While attacker effort was a limiting factor, probability-based prioritisation was a commonly accepted approach. But now attacker effort is exactly what AI is now reducing. 

If we look beyond what current information suggests is likely and towards what is possible, we start to see situations that could become a problem in future even if they are not one today.

The question is no longer ‘what's likely?’, and is now ‘what is possible and what could become possible as these tools evolve?’

Putting it to the test

One useful way to do this is with thought-based exercises, such as workshops and tabletop exercises, which helps teams to test ideas under pressure, improve decision making and response times, and work collaboratively to refine processes.

Researchers recommend that tabletop exercises and technical playbooks be updated to include scenarios where attackers rapidly adapt exploits across device models, abuse exposed management services, move laterally through engineering environments or cause device instability through failed exploitation. 

Running a tabletop exercise that determines how priorities are actually identified in your organisation will very quickly reveal if people are relying on probability or what is known to be possible. If the answer turns out to be ‘whatever the tool ranked highest’, that’s useful to know and definitely a cheaper one to discover before it reaches production.

From firefighting to prevention

None of this means abandoning AI tools altogether, because they have a place on the frontline. But the organisations best placed to handle AI-enabled attacks won't be the ones trying to out-race attackers. They'll be the ones who prioritise prevention by: 

• Reducing their attack surface.
• Reducing the rate of new vulnerabilities by catching them before they reach production.
• Testing how they make decisions before they have to make them under pressure.

TL;DR? Solid security fundamentals. You don't lose battles you don't have to fight.

If you’re ready to move from firefighting to prevention, let’s talk.