New findings reveal a security problem that should make every business leader stop and look at their website: Optimizely was found serving malicious JavaScript to the sites of 178 organizations, including Zoom, CNN, HSBC, Salesforce, GitLab, Ledger, BBC, Equifax, IKEA, and more. That list spans media, banking, software, crypto, retail, and credit reporting. It is not a niche website issue. It is a trust issue that reaches customers, brands, and the software supply chain.
The incident matters far beyond one vendor. It shows how a single third-party script service can become a delivery path for malicious code across hundreds of brands at once. And it arrives at the exact moment AI agents, AI browsers, and AI coding tools are being connected to the web. The future of AI depends on trustworthy software. This incident shows how fragile that foundation still is.
Security researchers found that Optimizely was serving malicious JavaScript to sites belonging to 178 organizations. The affected organizations include some of the most recognized names on the web: Zoom, CNN, HSBC, Salesforce, GitLab, Ledger, BBC, Equifax, IKEA, and more. The scale is the story. When a third-party service is embedded across many sites, a compromise does not stay contained to one company. It spreads through the trust that websites place in that service.
This is often called a client-side supply chain problem. Websites load code from outside vendors for analytics, experimentation, personalization, customer support, payments, and more. That code runs in the user’s browser with the same access as the site’s own code. If it is malicious, it can read page content, capture form data, track users, make network requests, or redirect behavior. Users do not see the vendor. They see the brand. That is why this kind of incident is so dangerous.
The 178-organization figure also highlights concentration risk. The modern web is built on shared services. That makes teams faster and cheaper to run. But it also means one weak link can create a blast radius that no single security team can see on its own.
Many organizations still treat third-party scripts as a marketing or analytics detail. That view is out of date. A script tag is executable code. It can change what a customer sees, what data leaves the page, and what actions are possible. In a banking site, that can mean account data. In a retail site, checkout behavior. In a software company, session tokens and user content. In a media site, reader privacy and ad delivery.
The hard truth is that most companies do not have a complete, live inventory of every script running on their own properties. Marketing teams add tags. Product teams add experiments. Agencies add pixels. Vendors add sub-vendors. Over time, the page becomes a chain of dependencies that no one fully owns. An incident like this one exposes that gap.
It also changes the vendor conversation. A vendor’s security questionnaire is not enough if the vendor can push new code to your site at any time. Businesses need to ask harder questions: What code is being served? Who can change it? How quickly can it be disabled? What monitoring exists for unexpected changes? What happens when a vendor is compromised?
The AI dimension here is not that AI caused this incident. The AI dimension is that AI systems are being built on top of the same web. AI browsers, shopping agents, research assistants, and task-performing “AI workers” are designed to visit websites, read pages, click buttons, fill forms, and sometimes make purchases or move data. If a site is serving malicious JavaScript, an AI agent may execute it, be manipulated by it, or act on false information without a human noticing.
That risk grows with autonomy. A human might pause at a strange pop-up or an unexpected request. An AI agent may follow instructions at machine speed and at scale. If an agent is logged into business tools, email, or payment systems, the impact can be severe. The same third-party script that targets a human visitor can become an entry point for an automated worker.
AI coding tools add another layer. Developers increasingly use AI assistants to suggest snippets, integrations, and third-party libraries. If those tools recommend a script without checking provenance or permissions, teams can add risk faster than they can review it. AI can help write code, but it does not automatically know your trust boundaries.
The lesson is clear: AI agents must treat the web as untrusted by default. They need sandboxes, allowlists, action limits, human approval for sensitive steps, and detailed logs. An AI worker that can click “buy” or “send” is a new kind of insider. It must be governed like one.
For business leaders, this incident is a stress test of digital trust. Customers expect a bank’s website to be a bank’s website. They do not expect a third-party service to be the weak point. When that trust breaks, the brand pays the price, even if the vendor made the mistake.
There are also contract and regulatory implications. Organizations should review vendor agreements for security notifications, subprocessor disclosure, audit rights, and the right to disable code immediately. Incident response plans should include a “third-party script compromise” scenario, not just a server breach scenario. Legal, security, marketing, and product teams need to practice together.
The most important shift is ownership. Someone in the organization must own the full lifecycle of third-party scripts: approval, deployment, monitoring, and removal. That owner needs authority across departments. A script approval process without enforcement is just a suggestion.
AI will be part of the defense. Security teams can use AI to analyze script behavior, summarize CSP reports, detect anomalies across many sites, and map which vendors touch which pages. AI can help overstretched teams see patterns that manual review would miss. It can also help draft incident communications and prioritize which scripts to disable first.
But AI will also be part of the attack surface. Attackers can use AI to obfuscate code, generate convincing phishing pages, find vulnerable integrations, and scale campaigns. The defense is not “more AI” by itself. The defense is AI with clear permissions, human accountability, and verifiable supply chain data. Organizations need to know what their AI systems can reach and what they are allowed to do.
This is where the future of AI and the future of web security meet. AI agents will not just read the web. They will act on it. That means every script, every vendor, and every permission becomes part of the AI risk model.
The web’s current model for third-party code is often “load and hope.” A site adds a script, trusts the vendor, and moves on. This incident shows why that model does not scale to an AI-driven internet. When AI agents browse, buy, and transact, the cost of blind trust rises sharply.
The next phase needs to be built on three ideas: provenance, permission, and proof. Provenance means knowing who published the code and whether it changed. Permission means limiting what that code and any AI agent can access. Proof means having logs and monitoring that show what actually happened. These are not just technical features. They are business requirements.
Companies that treat third-party scripts as invisible will face a larger blast radius as automation grows. Companies that build inventory, least privilege, monitoring, and AI guardrails will be able to move faster with less fear. The 178-organization list is a warning, not a final score. The real question is whether your organization can prove what runs on its site, and what your AI systems are allowed to do with it.
The future of AI will be powerful, but it will only be as trustworthy as the software supply chain beneath it. Trust must be engineered, reviewed, and monitored. This incident is a clear signal that the work needs to start now.