Website Security Is Getting Harder. AI Is Part of the Reason

The Old Vulnerability Is Still Useful

AI does not have to discover a new weakness to help an attacker be successful.

Google’s Threat Intelligence Group reported in 2025 that government-backed attackers used Gemini for coding assistance, research on potential targets, and information about publicly disclosed vulnerabilities. Google described much of the observed activity as productivity improvement rather than entirely new capability, and said safeguards blocked some malicious requests.

That distinction matters. An attacker can benefit from help understanding a published vulnerability or troubleshooting code that already exists.

Meanwhile, the organization running the affected software may have perfectly understandable reasons for leaving it alone. An ecommerce extension still handles orders. Replacing an older public-service application requires procurement and testing. A nonprofit’s donation tool is working, and nobody wants to disturb it during a campaign.

The difficulty is that “still working” says little about website security.

A sound maintenance arrangement needs to resolve those decisions before they become urgent. Who checks security notices? Who determines whether the website is affected? Can an important fix move outside the normal update schedule?

The Cybersecurity and Infrastructure Security Agency (CISA) recommends keeping software current and using its Known Exploited Vulnerabilities Catalog to help prioritize remediation.

Updates should be tested. But testing needs an owner and a deadline. Otherwise, “we need to test it first” can become an indefinite postponement.

The same review should remove software that no longer serves a purpose. An unused plugin is another component to track and maintain. If the campaign ended two years ago, find out why its software is still installed.

A Message That Fits the Recipient’s Job

Consider a hypothetical retailer preparing for a weekend promotion. An employee receives an email saying the payment connection needs verification. It appears to come from the ecommerce provider and references the store by name. The employee has a reason to act quickly. A checkout problem would jeopardize the promotion, and sorting it out falls within the employee’s responsibilities.

That context does much of the persuading.

AI can help produce convincing language and tailor a message using publicly available information. The FBI has warned that criminals use generative AI to make fraud more believable and carry it out at greater scale.

A nonprofit employee might receive the same approach during a fundraising campaign. At a government office, the request could appear to concern the vendor supporting an application portal. The details change; the pressure to keep a service working remains.

Voice messages can add another layer. In May 2025, the FBI warned about a campaign using texts and AI-generated voice messages to impersonate senior U.S. officials. The attackers sought to build trust before gaining access to personal accounts. That was an account-targeting campaign, rather than a reported website breach.

Its relevance to website security is the person holding the keys. Someone controls the hosting account, domain registration, and administrator access. Deceiving that person may offer a route around the protections built into the site.

Employees need a verification procedure they can follow without debating whether a message “feels suspicious.” Open the provider’s dashboard from a saved bookmark. Check an unexpected request through an established support channel. Use known contact information.

Administrative accounts should also have multifactor authentication, preferably phishing-resistant authentication such as FIDO/WebAuthn security keys or passkeys where supported. CISA recommends phishing-resistant MFA as stronger protection against account compromise.

An employee can be careful and still make a mistake. The account should have another defense.

Be Careful What You Connect to Your Own AI

Organizations are putting AI assistants on their websites because they can be useful. Answering a question at midnight or helping a customer locate an order has obvious appeal. But connecting an assistant to private records changes the job.

A chatbot answering questions from public pages needs relatively little access. An assistant retrieving customer orders needs to verify which customer is asking. One that changes an application or initiates a refund needs additional authorization.

The Open Worldwide Application Security Project (OWASP) prompt injection, sensitive information disclosure, and excessive agency among the risks affecting applications built with large language models. Prompt injection can include instructions embedded in material an AI system reads that attempt to redirect its behavior.

The underlying application must enforce the permissions. A written instruction telling an assistant to protect customer privacy cannot establish which records a user is entitled to see.

This deserves attention during development, while access can still be designed around the task. Give the assistant only the data and functions it needs. Test attempts to retrieve someone else’s information or trigger an unauthorized action.

The same restraint improves website security elsewhere. A content editor rarely needs permission to install software. A former contractor should not retain administrator access simply because nobody has reviewed the account. Broad permissions make the consequences of a stolen login larger.

The Security Plan Has to Survive a Busy Week

An organization can agree with every recommendation here and still struggle to carry it out. The hosting provider maintains the server, while another company supports the CMS. An employee handles user accounts. Security alerts go to an inbox that several people can access, although nobody is specifically assigned to read them. Before buying another tool, go and sort out those responsibilities.

A web application firewall, rate limits, and bot controls can help reduce certain kinds of malicious traffic. They need configuration suited to the site’s legitimate visitors and should work alongside updates and strong authentication.

Monitoring needs a responder. Backups need restoration tests. After a compromise, the team must fix the weakness, remove unauthorized access, and replace affected credentials before resuming normal service. CISA’s guidance emphasizes monitoring, protected backups, and tested recovery procedures.

Those arrangements should reflect what the website does. Restoring yesterday’s database could lose today’s orders or donations. A government agency may need to keep essential public information available while an affected service is investigated.

New Target Can Help

New Target helps businesses, nonprofits, associations and government agencies address website security through development, managed hosting and ongoing maintenance. Our hosting and support services include patching, monitoring, backups, and testing WordPress, Drupal, and WooCommerce updates before production.

Depending on the engagement, that work can include reviewing outdated components, tightening permissions, configuring protective services, and planning recovery.

A useful first step is to ask whoever maintains your website to explain what would happen if an administrator account were compromised tomorrow. Who would notice? What could the account reach? How would the team contain the damage?

If the answers are uncertain, contact New Target to turn them into a practical security and maintenance plan. We can help.