Offensive AI is changing the rules for critical infrastructure security
© Shutterstock
Most cybersecurity programmes were not designed for the speed of AI-driven attacks. Here is how leaders can close the gap...
For critical infrastructure, offensive AI is more than just a new threat. It also changes the pace at which cyber defence must work. Defenders have less time to act, and security must become continuous, with threats and risks assessed all the time. Adapting to these changes has a cost. That is why AI in cyber defence should be judged by the measurable gain it brings.
What do we mean by offensive AI?
The term covers two families of capability. The first finds weaknesses in software and turns them into working exploits. Defenders can use the same capability to find and fix flaws before attackers do. The second carries out the attack itself, with agents that adapt as they go.
Agents of this kind have already found their way out of controlled test environments and taken unauthorised actions against real systems and organisations with no human attacker directing them. Several reports published since July 2026 describe these incidents, including disclosures from OpenAI, Anthropic and the UK AI Security Institute.
Both families draw on the most capable “frontier” models, as well as on “open weight” models, which anyone can download, run and modify.
First challenge: less time to defend
For defenders, the biggest change is the compression of time. Most security programmes assume there is room to assess, schedule and deploy a fix before anyone exploits a flaw. Yet VulnCheck found that in 2025, nearly 30% of vulnerabilities known to be exploited were attacked on or before the day they were made public. With AI, attackers can spot and exploit flaws much faster.
Proven security practices still apply, but they must now run at machine speed. European law has also moved to deadlines counted in hours. Since 11 September 2026, the Cyber Resilience Act has required manufacturers to send an early warning within 24 hours when a vulnerability in one of their products with digital elements is actively exploited. NIS 2 already requires organisations in critical sectors to alert the authorities within 24 hours of a significant incident.
Second challenge: from periodic to continuous
Quarterly scans, annual penetration tests and monthly patching were designed for a time when defenders had weeks to fix a flaw before attacks began. But today’s attackers do not wait for the next scan or patch cycle. An organisation that still relies on this periodic rhythm risks falling behind. The way forward is security that is continuous, automated and risk-driven.
Maintenance windows still work for some systems, but not every critical flaw can wait for the next one. Continuous assessment brings a new difficulty: prioritisation at scale. As finding flaws gets easier, the real work shifts to choosing which ones to fix first. Many teams still rank flaws by severity score alone.
A better guide is whether a flaw is already used in attacks and whether an attacker could reach it in your systems. By that measure, a moderate flaw on an exposed server can matter more than a critical one in a component that is switched off.
Critical infrastructure: the most difficult case
In critical infrastructure, ageing systems and software, often called technical debt, have become an operational risk. Most organisations respond to shorter timelines by patching faster. In this sector, that is often not possible. Their equipment stays in service for decades, certified systems may lose their certification if changed, maintenance windows are rare and some components have reached the end of their support. In industrial processes, updates must be tested with care, since a faulty one can stop production or affect safety. The problem is structural; it cannot be solved with more effort or more budget.
AI makes this debt harder to live with. Finding flaws in older systems and firmware once took rare skills and a lot of time. With AI, attackers can do much of that work, even without the source code. Operators that cannot patch must continuously track which weaknesses an attacker could reach.
Operators also depend on one another and on shared components, so a single compromise can spread well beyond its source. Geopolitical tension adds pressure from two sides. The EU cybersecurity agency ENISA reports that internet-exposed OT systems remain high-value targets for all types of attackers, so faster, cheaper attack tools matter here.
At the same time, cyber defence that relies on a single AI provider can be disrupted by export controls or policy changes, so operators need the freedom to use more than one model.
“The systems our customers run are built to last for decades and rigorously certified. Those are strengths, and we believe security should be designed around them.”
Christophe Bianco, Vice President, Thales Cybersecurity Services
Our view: five pillars for continuous defence
To keep pace with increasingly automated threats, organisations need to run cybersecurity as one continuous capability rather than as a collection of standalone controls.
At Thales, we structure this response around a cross-functional transformation that provides visibility and supports decision-making at every stage. This approach is supported by five complementary pillars. Across all five, AI-assisted defence has a running cost and is not an end in itself: its adoption must be tied to a measurable gain in risk reduction, processing time or operational capacity.
1. Secure applications
As AI accelerates software development and expands the attack surface, organisations must strengthen the security of both existing and newly developed applications. This means embedding security from design through to deployment while industrialising remediation processes.
2. Detect and respond at machine speed
The time between compromise and impact has shrunk. Security teams must therefore be able to investigate, prioritise and contain threats at machine speed, while humans supervise the process and focus on critical decisions, complex investigations and strategic oversight.
3. Connect security operations in a unified platform
Detection, vulnerability management, threat intelligence and exposure management can no longer operate in isolation. Brought together in one orchestrated operating model and supported by AI agents, they allow organisations to find, assess and fix risks faster.
4. Test and assess continuously
Organisations need a current view of their exposure and regular tests showing that their defences still work, so that security spending matches the threats they actually face.
5. Protect what cannot be fixed
When a system cannot be modernised, patched or replaced, organisations should reduce its exposure, isolate it where they can and surround it with compensating controls. A Zero Trust approach helps contain an intrusion so that essential services keep running.
Where decision-makers can start
Automate cyber defence
Offensive AI compresses attack timelines, so defence must speed up too. Automation can accelerate detection, investigation and response while human judgement is kept for the decisions that carry the greatest impact.
Tackle technical debt as a resilience issue
Legacy environments are an operational risk as well as an IT challenge, because they limit how quickly an organisation can respond to threats. Systems that cannot be modernised should be identified, prioritised and, where necessary, isolated.
Preserve strategic autonomy
Managing dependencies is now a priority in its own right. Decisions about AI models, hosting environments and technology providers should be driven by business criticality, regulatory requirements and the need to ensure long-term resilience and autonomy.
None of this depends on a single technology. Moving to continuous defence is a question of how security is organised, governed and funded, and the first decisions belong to leadership.
For critical infrastructure, continuous defence is not about moving faster at any cost. It is about matching the speed of the threat while preserving the stability, safety and resilience on which essential services depend.
Thales sets out this approach in more detail in Deciding faster than the threat. This paper also covers how to protect systems that cannot be patched.