October 22, 2026

expert reaction to the PHP cybersecurity issue

Scientists react to the PHP cybersecurity issue that led to website being taken offline worldwide.

Prof Raj Rajarajan, Director of the Institute of Cyber Security at City St George’s, University of London, and co-founder of Cyber London, said:

“PHP has been implemented since 1995 as part of backend servers. A vulnerability was recently found and a patch was developed. However, the news got leaked quickly, hence there has been a rush to apply the patch quickly as all webpages rely heavily on PHP.

“In the age of AI we are going to see more of these types of attacks as threat intelligence software have multiple AI agents for threat hunting. These will be working in the background and will penetrate systems where they identify a backdoor.

“The challenge will be to protect legacy technologies as many organisations may not know all their in-house technologies. Even if they know, they may not have the people with skills to address those legacy technological challenges.

“The speed at which AI transformation is happening is worrying and we need to have good guardrails in place to stop these kinds of vulnerabilities escaping and getting into cyber criminals’ hands.”

Steve Sands, a Fellow of BCS, The Chartered Institute for IT, said:

“There are likely to be many more unknown vulnerabilities in software that AI tools will find in the coming months.  Organisations need to scale up their ability to respond to these in a similar fashion to this one.  This will certainly be more challenging for smaller organisations and lots of the big companies have patched this quickly . But even for larger ones it shows that information security professionals need to be members of leadership teams including on company boards, and that is not always the case.

“This could be compared to a critical safety recall in the automotive sector affecting millions of cars.  The sector isn’t presently geared up to fix millions of cars in a single day or week.  So car owners are left with a dilemma – do I follow guidance and stop using my car until I can be scheduled for the fix, or do I take the risk?  The difference is that the risk is no greater for the car driver today than it was yesterday, before the safety recall was announced, but the risk has changed for those running vulnerable PHP as the vulnerability is now known by actors who want to abuse it.

“Taking systems offline is for many organisations a very last resort.  Their online presence is often linked to their revenue streams.  So anything which disrupts availability has significant financial implications.  

“Also, many organisations don’t have the necessary records showing exactly what software runs in which systems.  There has been a growing recognition of the importance of a Software Bill of Materials (SBOM), which documents all software in a system.  As more vulnerabilities are now being detected (by the good guys and the bad guys), it’s even more important that organisations can identify (quickly) whether they have the affected software and be able to react.”

Dr Shahzad Memon, Senior Lecturer in Computer Science and Digital Technology at the University of East London, said:

“Taking a website offline can look like a security failure or a cyberattack, but it may be the decision that prevents a successful attack. Where organisations cannot immediately apply and verify security updates, temporarily withdrawing an affected service can reduce its exposure to attack.

“The two critical PHP vulnerabilities are serious because they could allow attackers to run code on affected servers without logging in. Depending on the system and the access gained, that could put sensitive information and services at risk.

“For the public, an unavailable website does not automatically mean it has been hacked. Equally, a precautionary shutdown cannot guarantee that no information has been compromised. Organisations need to explain what they know, what they are checking and when users can expect another update.

“The priority is to apply the fixes, check for signs of compromise and restore services securely. A temporary outage is disruptive, but bringing a vulnerable service back too quickly could have far greater consequences.”

Dr Soteris Demetriou, Associate Professor in the Department of Computing at Imperial College London, said:

“The PHP2Shell vulnerabilities are alarming, but not surprising. It has been clear for some time that AI makes vulnerability discovery faster and cheaper. The flawed code had been present in PHP for more than ten years, and frontier models made it practical to understand the weakness and build a working exploit within days. In this case (white hackers or defenders) found the problem first and coordinated a fix with the PHP Foundation which is the good news, although this does not mean this had not been or couldn’t have been independently discovered by malicious actors. For this incident difficult part begins now. The patch has to reach millions of servers, many of which are run by small teams or are no longer actively maintained, while attackers with access to similar models can analyse the same patch to identify systems that remain exposed.

“The incident illustrates the gap between how quickly vulnerabilities can be found and how slowly they are fixed. In most organisations, the path from red teaming or incident response to patching, testing and deployment still depends on manual processes that take weeks or months. Agentic AI needs to become part of these everyday workflows, so that remediation and deployment can keep pace with the speed at which adversaries can now discover and exploit vulnerabilities.

“Giving AI agents this role is tempting but raises its own problem. A patch generated and deployed in minutes is useful only if we can verify that it addresses the actual flaw and introduces no new ones. For this reason I see faithful explainability as a security requirement. An explanation of an agent’s actions or output is faithful when it reflects the reasoning that actually produced the outcome. A plausible justification constructed afterwards does not meet this standard. In our research group at Imperial (https://apss.doc.ic.ac.uk/) we design state of the art methods to produce explanations of this kind that a human expert or another system can trust, inspect and interact with. We are currently extending this approach to the execution traces of AI agents, so that individual steps taken by an agent can be audited. Organisations will be reluctant to give agents authority over production systems unless accountability improves at the same rate as speed.

“The incident also underlines the responsibility of model providers and of the organisations that deploy their models. Both need to make misuse difficult through rigorous safety alignment or retraining, guardrails, and system-level safeguards. These safeguards are more effective when we can explain why a model refused or complied with a request, because it then becomes easier to detect when a safeguard has been circumvented. Our recent work on concept unlearning, to be presented at a world-leading cybersecurity conference (ACM CCS) in November, studies how harmful capabilities can be removed from a model or unlearned in a practical way that can be verified which is a concrete measure against model jailbreaking attacks which aim to bypass its safeguards and misuse it. We should nevertheless be realistic about the limits of such measures. Knowledge can still be distilled from non-safety aligned models, surrogate models can be trained with safeguards removed. Each additional layer of protection still increases the cost and expertise an attacker needs.

“In summary, we need to be composed here and go back to system security fundamentals. We should follow a defence in depth approach, privilege separation within systems to decrease the blast radius of exploits and complete mediation. We should raise the cost of attack, adapt our processes to the speed that AI now makes possible, require that the AI systems we depend on are explainable and auditable, and accept that responsibility is shared by everyone in the chain, from model developers to the organisations that deploy them, regulators and standards bodies alike.”

Dr Saira Ghafur, Director for Digital Health, Institute of Global Health Innovation, Imperial College London, said:

“The PHP vulnerability highlights that healthcare’s biggest cyber risks often sit in everyday web software, not just clinical systems. Patient portals, booking tools and supplier platforms are frequently built on PHP, and when a flaw this severe is already being exploited, the window to patch is reduced to hours, not weeks as the consequences can be far reaching. The NHS has moved quickly to alert its partners as this vulnerability needs urgent action, but the real test is the multitude of third-party suppliers and smaller sites. Organisations need to patch now, check their suppliers have done the same, and have a plan to isolate exposed systems without significant disruption to patient care.”

Martyn Thomas FREng, software engineer and founder of software engineering company Praxis, said:

“We have known for decades that most programmers make an error in about every 30 lines of programming and that more than half of them are not corrected before the software is released. There are undoubtedly huge numbers of errors in most systems now in use although most of these will cause a problem only under very rare circumstances. 

“The arrival of advanced AI turns these errors into much higher risks. We can expect an increase in cyber attacks as AI systems are used more widely to search for exploitable faults. 

“It makes sense to keep important data and control systems off the open internet wherever possible.”

Dr Fazl Barez, Senior Researcher and Lead of the Oxford Martin AI Governance Initiative Safety Lab at the University of Oxford, said:

“Cyber is particularly at risk of AI because attacks don’t need physical access and can be done at a large scale. We’ll soon see more examples of AI-discovered vulnerabilities that human experts either missed or would have taken them a long time to find.

“The bigger and more important question is how we govern and control AI systems whose capabilities are growing faster than our ability to understand and secure them.”

Dr Daniel Gardham, Lecturer at the Surrey Centre for Cyber Security, University of Surrey, said:

“The discovery of these vulnerabilities illustrates how AI is changing the pace of cyber security. Flaws in software previously may have remained hidden for years because discovering them requires considerable specialist expertise and time. AI is reducing those barriers and making it easier for non-experts to develop and exploit attacks. 

“The vulnerability found in PHP is serious: it could lead to remote code execution, where attackers may be able to take control of affected systems. PHP underpins a substantial proportion of the web, so a critical vulnerability in its core will have consequences far beyond any single organisation. The decision to take so many major websites offline highlights the potential impact of these vulnerabilities.

“For affected organisations to remain safe, they need to know what software they are running, apply security updates promptly and investigate whether their systems have already been compromised.

“The difficulty is that AI gives attackers the same advantages it gives defenders. Finding vulnerabilities is becoming faster and easier, but fixing them across the huge range of systems that rely on this software is much harder. How well organisations manage that gap will determine whether AI makes the internet safer or more vulnerable.”

Quotes supplied by the all-Ireland SMC:

Dr Deepak Padmanabhan, Senior Lecturer, School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, comments: 

“PHP is a scripting language that is employed by websites. When you send a request to the website, such as querying a national-level data service about demographic make-up or voting statistics of a region, it would usually be a PHP script that queries the internal database and fetches the information, which then gets formatted in ways that it would appear in the response web page.

“The current security vulnerability is one which could potentially, under technically rare circumstances, allow somebody to run a program through remote access, one that is called remote code execution. In other words, malicious actors could exploit this to run *their own* program on the servers, leading to myriad outcomes. It could involve querying the underlying database in ways that are unintended or undesirable, or to run a deliberately computationally intensive program so the server is rendered incapable of routine function.

“This remains an extremely rare possibility and was intended to be fixed through some program updates. However, the information about this vulnerability leaked earlier than the roll out of updates, triggering the panic we are witnessing. The updates, as far as one can reliably learn, would close the vulnerability – thus, this is only a transient issue that exists until fixes get applied widely.

“Some news outlets have cited claims that Russian state actors or organised criminals could exploit the flaw to carry out cyber-attacks. However, at the time of this writing, there is not an iota of evidence to suggest that such threats are imminent.” 

Dr. Domhnall Carlin, School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, comments: 

“PHP is a programming language from the mid-1990s, which runs behind an estimated 70% of websites and web applications.  It also drives many systems that allow users to build interactive websites, like WordPress, Drupal etc.So, when a vulnerability is detected in the underlying programming language, the blast radius can be very broad. 

“Two critical vulnerabilities have recently been disclosed in PHP, now referred to as “php2shell”. These have been given a CVSS score of 9.2 out of a possible 10, meaning they are of critical severity.  They allow for unauthenticated remote code execution in PHP and affect all PHP installations serving web requests – a huge attack surface across the globe. 

“The good news is that a fix is available.  However, updating can be a slow process because PHP is often buried several layers deep within the tech stack. Many organisations also lack a complete inventory of where vulnerable versions are running. Even when this is known, applying patches or updating PHP versions can also break custom code and plugins, so fixes need substantial testing, which sometimes results in downtime. Limited resources make this harder still and smaller sites may depend on hosting providers or developers to act on their behalf. 

“There seems to have been a huge co-ordination process behind the scenes to resolve this issue before it became public.  The remediation advice at https://www.php.net/ should be followed. If patching can’t happen immediately, the NCSC advises taking affected systems offline until they’re mitigated.” 

Declared interests

Steve Sands: “No conflicts of interest”

Shahzad Memon: “I have nothing to disclose.”

Saira Ghafur: “No COI”

Dr. Domhnall Carlin: “nothing to declare”

Dr Deepak Padmanabhan: “I have no interests to declare”

For all other experts, no reply to our request for DOIs was received.