How AI is changing the software your small business relies on

What the PHP2Shell security flaws tell us about software security, custom tools and the growing role of AI agents in everyday business.

Research checked: 23 October 2026

You can decide against buying an AI tool and still find AI changing the software your business depends on. Your hosting provider may face newly discovered security weaknesses. Your developer may use AI to build features. Your CRM supplier may introduce an assistant that can update customer records.

The PHP2Shell vulnerabilities disclosed in October 2026 bring one part of that change into focus. Researchers used AI to help investigate serious flaws in established software, prompting urgent security updates.

What happened with PHP2Shell?

On 21 October 2026, PHP released security updates addressing two critical vulnerabilities: CVE-2026-109618 and CVE-2026-109619. Under certain conditions, an attacker could use these flaws to execute code on a server without logging in. The fixed releases included PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, according to the PHP security advisory.

PHP is software used to run websites and web applications. You may never have chosen it yourself, but it sits underneath platforms including WordPress.

The vulnerabilities were in PHP itself. WordPress's security announcement explained that updating WordPress would not resolve them: the underlying PHP installation needed attention. A website could therefore need urgent work even if its owner had kept WordPress up to date.

AI helped researchers understand and demonstrate the weaknesses. Wiz's account of its investigation describes using AI to develop working exploits and adapt them to different configurations. This was AI-assisted research into existing flaws, rather than evidence that AI-generated code had caused them.

There were also attack attempts. Amazon Threat Intelligence detected an attempt against its honeypot systems on 17 October, before public disclosure. Amazon cautioned that it was unclear whether the attempts would have succeeded against real-world systems. That distinction matters: an attempted attack is not proof that a business was compromised.

If your business runs a PHP-based website, ask your host or developer to confirm that the relevant security fix has been applied. The NCSC's PHP guidance sets out the response priorities.

How will AI change business software?

Three developments are worth separating because each affects a different business decision.

First, AI can help researchers scrutinise existing software. PHP2Shell illustrates how tools can make complex weaknesses easier to investigate. This creates opportunities for defenders alongside risks from attackers.

Second, AI coding assistance can help create or modify software. A business might commission a focused internal tool where previously it would have used a spreadsheet or accepted an awkward manual process.

Third, AI agents can take actions through connected systems. Microsoft's Copilot Studio documentation already describes agents that respond to events and execute tasks within defined permissions.

These developments overlap, but they are not interchangeable. Finding a security flaw, writing an application and deciding to send a customer an email involve different capabilities and controls.

Our recommendation is to treat this as a question of software ownership. Before expanding what a system can do, establish who checks its work and who can step in when something fails.

1. Check who handles urgent software updates

The PHP incident shows why maintaining a website involves more than updating the parts visible in an administration panel.

A small business may deal with an agency, a hosting company and several software suppliers. When a vulnerability appears, uncertainty about responsibility can delay the response.

The NCSC's assessment of AI's impact on cyber threats through to 2027 expects AI to make elements of cyber intrusion more efficient and effective. It identifies vulnerability research and exploit development as particularly significant areas. This is an assessment of likely developments, rather than a prediction that every website will be attacked successfully.

Ask your provider which layers they maintain, how they monitor urgent advisories and how they confirm that a fix has been applied. For an important business system, agree who decides whether it should temporarily go offline.

You do not need to become the technical specialist. You need a named person or supplier responsible for the response.

2. Judge custom tools on their running costs

AI-assisted development makes it reasonable to explore smaller, more specific tools: a job handover tracker, a reporting dashboard or a quoting interface built around an established price list.

However, evidence about development speed depends heavily on the task. In GitHub's 2022 controlled experiment, developers using Copilot completed a specified JavaScript task about 55% faster. That finding does not mean an entire software project becomes 55% cheaper.

METR's early-2025 study found that experienced developers working in familiar open-source projects took 19% longer when AI tools were allowed. Its February 2026 update suggested newer tools likely offered greater benefits, while explaining why its follow-up data could not reliably establish the size of that improvement.

For a business buyer, the sensible response is to test the economics of the proposed tool. Include hosting, support, integrations and future changes alongside the initial build.

Suppose you want an internal quoting tool. A prototype might look convincing, but someone still needs to check calculations, access permissions and how it handles missing information. You also need a way to update prices without rebuilding the application.

Consider adapting your existing software first. Commission a bespoke tool when the fit is meaningfully better and someone can maintain it. Our AI automation services explain how focused tools can fit into existing processes.

3. Give AI a narrow job before granting wider access

An AI assistant that drafts a reply has limited authority. An agent connected to your inbox and customer database may be able to change records or communicate with customers.

That distinction is useful when comparing products. Our guide to AI chatbots and AI agents explains the difference in more detail.

Consider a possible workflow for a small service business. An enquiry arrives by email. AI extracts the requested work and prepares a summary. A normal calculation uses approved prices to create a draft quote. A member of staff checks it before sending.

This example uses AI where interpreting varied language is helpful. The price calculation follows fixed rules. A person approves the commercial commitment.

An initial version might only prepare summaries, without changing customer records. Once it performs reliably, you could allow it to create draft entries. Microsoft recommends a similarly gradual approach to autonomous agents, with restricted permissions and human oversight for sensitive actions.

Choose the smallest amount of access that lets the tool do its job. If it only needs to read a price list, it should not be able to edit it.

4. Test mistakes and unusual inputs

A demonstration usually shows what happens when everything goes well. Your evaluation should also cover unclear requests, duplicate messages and unavailable systems.

For the quoting example, try an enquiry with no quantities, a customer asking for an unauthorised discount and an email that contradicts an earlier message. Decide when the workflow should stop and ask someone to review it.

There is also a security issue specific to language-based AI. The NCSC explains that prompt injection can occur when a model treats untrusted content as instructions. For example, malicious directions embedded in a document might influence an assistant processing that document.

The possible damage depends partly on the system's permissions. A tool that drafts an internal summary presents different consequences from one authorised to send files externally.

Ask suppliers to show how they limit actions outside the model, record what happened and handle failed integrations. Clear instructions to an AI are useful, but access controls and approval steps need to support them.

5. Budget for maintenance and recovery

Faster software development does not remove the work needed after launch.

The PHP Foundation's account of the security release explains that experienced engineers still had to weigh compatibility issues and review fixes. It also describes growing volumes of AI-assisted vulnerability reports, which people need to assess.

The lesson for a business commissioning software is practical: agree ongoing ownership before the tool becomes essential.

Who notices if an automation stops running? Who renews a connection when authentication expires? Can another supplier understand the setup? How would you restore data after an incorrect update?

Recovery also requires care. Restoring a database cannot retract an email already delivered to a customer. Some actions deserve approval precisely because their effects extend beyond the system you control.

Keep a short operating record covering the tool's purpose, owner, connections and recovery process. It should help someone take over without reconstructing the project from old messages.

What could business software look like next?

A plausible direction is that more work begins with a request, while software coordinates the steps behind it. Instead of opening several applications, a business owner might ask for a draft proposal based on an enquiry and an approved service list.

Existing agent capabilities support that possibility. They do not establish that every business can delegate whole processes reliably or affordably.

Another possibility is more small tools built around particular businesses, connected to established accounting or customer systems. That could reduce awkward workarounds without requiring a wholesale replacement of trusted software.

Both developments make integration quality and maintenance important purchasing criteria. A useful tool should save enough effort to justify the checking and support it requires.

Start with one repeated task whose output is easy to assess. Measure the time spent reviewing and correcting it as well as the time saved. Expand its responsibilities when the results justify it.

Frequently asked questions

Did AI cause the PHP2Shell vulnerabilities?

The published research describes AI helping investigate and demonstrate weaknesses already present in PHP. It does not establish that AI-written code caused those vulnerabilities. AI's role included helping researchers develop and adapt working exploits.

Can updating WordPress fix PHP2Shell?

Updating WordPress alone does not fix these vulnerabilities because they affect PHP itself. Ask your hosting provider or developer to confirm that the underlying installation has received the appropriate PHP security update or vendor-provided fix.

Should a small business replace paid software with AI-built tools?

Consider a custom tool when it solves a specific problem better and its ongoing costs are justified. Compare it with configuring your current software. Include testing, support and maintenance in the decision, particularly if the tool handles customer data or important transactions.

Which actions should an AI agent need approval for?

Prioritise approval for actions with financial consequences, external communications, sensitive data access or changes that are difficult to reverse. The right boundary depends on the task and its impact. Begin with drafts or recommendations where possible.

Choose one process to review

Ask your website provider about PHP updates if relevant, then look at one business process where software creates unnecessary manual work. Establish what a better version would do, what access it would need and who would maintain it.

If you would like help assessing that opportunity, our AI consultancy for small businesses can help you decide what is worth building. Tell us about the process you want to improve.