The current PHP security crisis illustrates a defining challenge for cyber resilience: artificial intelligence can expose weaknesses that have survived years of scrutiny, while making those same weaknesses easier to exploit. The immediate priority is protecting affected systems. The wider lesson concerns how quickly organizations can turn new defensive capabilities into practical protection.
On 21 October, PHP released emergency updates addressing two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619. These can allow attackers to execute code remotely without authentication. PHP underpins widely used web platforms including WordPress, Drupal, and Joomla, making weaknesses in its core engine a concern far beyond any single website.
Researchers at cyber security specialist Wiz used frontier AI models to uncover and demonstrate serious flaws in code that had existed for more than a decade. AI helped with the difficult work of combining weaknesses, testing exploit paths, and adapting them to different applications. According to Wiz, this substantially reduced the specialist effort needed to turn complex memory problems into working exploits.
There is an important qualification to the idea that these vulnerabilities were entirely hidden. The PHP Foundation says the underlying problem behind one vulnerability had been identified in 2024. Remediation stalled because of technical trade-offs, limited resources, and difficulties reaching decisions. Subsequent research established the practical danger. AI helped expose the full significance of a longstanding weakness that had remained unresolved.
This is the positive side of the AI coin. Used by defenders, the technology can help uncover dangerous defects, establish their consequences, and validate remedies. Finding a vulnerability creates an opportunity to remove it before criminals cause widespread harm. Leaving it undiscovered offers no assurance of safety.
The other side of the coin emerged as exploitation attempts targeted vulnerable systems. Amazon Threat Intelligence reported attempts on 21 October before the patches were released and identified an earlier attempt dating to 17 October. These were attempts to exploit vulnerabilities before a public fix was available: a genuine zero-day exposure. Amazon cautioned that it was unclear whether those early attempts would have succeeded against real systems.
While the available evidence does not confirm that attackers used AI in these particular attempts, it seems highly likely. Amazon itself explicitly warns that AI-assisted analysis makes it easier to turn patches or advisories into working exploits. The defensive research itself demonstrates how much AI can accelerate that process.
The central problem is not the ever-increasing capabilities of AI; it is the stock of historical vulnerabilities that may be embedded in software on which organizations depend. The current PHP crisis is unlikely to be an outlier – it is more likely to be a forerunner of issues to come in other areas. It will take time to surface vulnerabilities created years ago when coding and testing took place in a totally different environment.
AI is a capability that changes who can discover and exploit those weaknesses and at what speed. It did not create these PHP defects. But it has changed the practical consequences of leaving them unresolved.
For defenders, the implication is that AI adoption must translate into faster identification, validation, and remediation. Implementing an AI-enhanced security product achieves little if discoveries then wait in queues, responsibility is unclear, or emergency changes cannot be authorised. Technical capability needs operational capacity behind it.
The immediate official advice is clear. PHP urges users to install fixed releases, while CISA calls for immediate patching. ENISA stresses identifying PHP applications, coordinating with providers, and prioritising Internet-facing systems. The UK NCSC also highlights software that bundles PHP and advises considering temporary withdrawal of systems that cannot be adequately protected.
For resilience professionals, this crisis confirms yet again that dependency visibility and response readiness is essential. Questions to be asking include: Which critical services depend on PHP, directly or through suppliers? Who can confirm that fixes are deployed? What happens if an exposed service must be disconnected? How will customers be supported while normal service delivery is interrupted?
Organizations should also use this incident to revisit weaknesses previously judged too difficult to exploit, especially in legacy software created many years ago. AI may invalidate that judgement without any change to the underlying software. Assurance therefore needs to consider evolving attacker capability, alongside the current state of technical controls.
That review must connect security findings to business decisions. Leaders should establish who can accept temporary disruption, fund urgent remediation, and challenge suppliers when assurances are incomplete. Exercises can test whether these decisions happen quickly enough when several critical dependencies require attention simultaneously.
Both sides of the AI coin lead to the same imperative: defenders must be as ready to use AI for protection as attackers are to use it for exploitation. The opportunity is to reduce accumulated vulnerability before it becomes disruption. Success depends on combining faster discovery with timely decisions, maintained software, and effective action.
The author
David Honour is editor of Resilience Forward.
ChatGPT and Gemini helped provide background information for this article.
