AI is getting better at finding security flaws. Can your business keep up?

When a serious security flaw makes the news, the first question is usually whether your business is affected. The next is how quickly you can do something about it. AI adds another question: what happens if finding dangerous weaknesses becomes faster than fixing them?
The recent PHP2Shell disclosures bring that into focus. They show how AI-assisted research can help establish that a software problem has serious security consequences. They also expose a less glamorous challenge: turning a discovery into protection across real systems, suppliers and businesses.
For business leaders, the lesson isn't that AI has made cyber security hopeless. It's that knowing what you run, who owns it and how to update it matters even more when vulnerability research accelerates.
What actually happened with PHP?
On 21 October 2026, PHP disclosed and patched two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619, collectively known as PHP2Shell. Under certain conditions, they allow an attacker to execute code on an affected server without first logging in. Both carry a CVSS v4.0 severity score of 9.2, according to the PHP Foundation's account.
PHP is software used to run web applications. That makes this relevant beyond businesses whose own developers write PHP: an organisation can depend on it through a hosted website, third-party application or supplier-managed service.
The discovery story needs some care. The PHP Foundation says developers identified early signs of the underlying out-of-memory problem in 2024, following a Quarkslab audit. Researchers demonstrated its remote-code-execution implications in August 2026. OpenAI's account says Calif used its GPT-5.6 Sol model in the research, while Wiz independently discovered and reported both vulnerabilities.
So this isn't simply a story about AI finding a completely unknown bug. The more useful interpretation is that AI-assisted research helped establish how dangerous an existing technical problem could be.
There is also a separate follow-up. On 23 October, PHP disclosed CVE-2026-115560, a related remote-code-execution vulnerability not addressed by the original updates. The Canadian Centre for Cyber Security's updated advisory describes an uncommon combination of configuration and application behaviour needed for exploitation. The evidence reviewed for this article does not establish that this additional flaw was AI-discovered.
This article reflects reporting and guidance checked on 24 October 2026. Remediation advice may change as further updates become available.
The future risk isn't just more bugs
Frontier models, the most capable general-purpose AI models being developed, could change the effort required to investigate software weaknesses. The important possibility isn't just a larger pile of findings. It's that problems previously considered difficult to exploit may become easier to understand and turn into practical attacks.
The PHP timeline illustrates that distinction. Recognising that software behaves incorrectly is one achievement. Establishing that the behaviour can let someone take control of a server is another. If models help researchers bridge that gap, organisations may need to reassess unresolved issues sooner than expected.
That doesn't prove every future model will reliably produce working attacks. Nor does it mean every critical vulnerability becomes immediately exploitable everywhere. In its 23 October analysis, CrowdStrike said reliable PHP2Shell exploitation required adaptation to particular deployments. It reported probing and unsuccessful attempts, but no successful exploitation within its visibility. That is a limited observation, not evidence that nobody had been compromised.
The sensible planning assumption is that attackers may need less time or specialist effort for some parts of their work. Businesses shouldn't build their security around the hope that exploitation will remain too difficult.
Defenders get the same opportunity, but not automatically
More capable models can also help legitimate researchers investigate weaknesses before attackers exploit them. OpenAI's disclosure account describes both the defensive opportunity and the risk that similar capabilities support offensive tools.
But discovery is only one stage. Someone still has to validate the finding, develop a compatible fix, distribute it and deploy it without breaking essential services. Faster research doesn't automatically accelerate that whole chain.
It can also create noise. The PHP Foundation reports increased volumes of invalid, duplicate and unreproducible submissions alongside AI-assisted research. More reports aren't necessarily more useful security knowledge. Human validation remains essential, particularly where a poor recommendation could interrupt a business-critical application.
The future advantage may therefore belong less to whoever generates the most alerts and more to whoever can reliably turn credible findings into completed fixes.
Start with exposure, not an AI shopping list
For a business responding now, the first task is to establish where PHP is used. Ask internal IT, hosting companies and application suppliers to check, rather than assuming a website content update covers the underlying software. ENISA's guidance explicitly includes third-party applications and products.
Prioritise internet-facing services, then confirm the affected runtime and the appropriate supplier-supported fix. The Canadian advisory lists PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12 as the original fixed releases, alongside applicable vendor-backported packages. Those fixes must not be treated as resolving the separate 23 October vulnerability.
For that follow-up issue, the advisory's interim guidance is to disable session upload progress while awaiting an update. Your technical team should check the latest guidance and application implications before making changes. This mitigation is not a substitute for patching the original flaws.
Ask for evidence of what is actually running after remediation. An update being available, approved or downloaded isn't the same as protection being deployed. Where compromise is suspected, investigate separately: patching closes a weakness, but doesn't establish whether it was used beforehand.
Make the response repeatable
The longer-term improvement is to remove uncertainty before the next urgent disclosure. Keep a usable record of critical services, their dependencies and their owners. Agree who can authorise emergency changes and how suppliers will confirm completion.
Speed still needs judgement. An untested update can interrupt operations, but an ordinary change schedule may be too slow for an exposed critical system. Plan proportionate testing, recovery options and escalation routes in advance, so each incident doesn't trigger a fresh debate about responsibility.
For businesses working with an MSP alongside internal IT and application suppliers, there should be one agreed view of outstanding exposure. Ownership shouldn't disappear between organisations.
PHP2Shell doesn't establish that AI will inevitably favour attackers or defenders. It does show why discovery and response must be considered together. As frontier models improve, the practical goal is straightforward: shorten the distance between a credible warning and a verified fix.
