Within three days, 178 major websites were compromised through the same class of hijacking technique. Java servers were controlled by a Base64 string, while an Nginx flaw that had remained hidden for 10 years may affect roughly 225 million websites. What stands out is that none of these incidents relied on a new AI-powered attack tool. Instead, they emerged just as AI is accelerating both software production and migration.

What this is

Fireship compiled three security incidents. The first involved Optimizely, where attackers hijacked websites through its frontend script modification capability. The second exploited a Java deserialization vulnerability, allowing attackers to control servers with Base64-encoded content. The third was an Nginx vulnerability that had existed for roughly 10 years and could affect as many as 225 million websites.

Paper.design took a different product approach: letting coding agents directly manipulate real HTML and CSS pages rather than merely generating code as text. Coding agents—AI systems capable of autonomously breaking down tasks, invoking tools, and modifying software—thus gained a workspace much closer to web design.

Industry view

We believe these three incidents point to the same conclusion: AI is increasing the speed at which software is produced, but the security debt embedded in legacy systems will not disappear merely because code output rises. The faster companies allow agents to scan codebases in bulk, modify configurations, and connect to production environments, the greater the opportunity for misconfigurations and historical vulnerabilities to be exploited at scale.

The risk is not confined to the attack side. Tools such as Paper.design move AI from “suggesting code” to “directly modifying pages,” improving design and development efficiency while also expanding privilege boundaries. If agents can access live websites, component libraries, or deployment systems, their erroneous changes can immediately affect production operations.

Impact on regular people

For enterprise IT: System inventories, dependency updates, and privilege audits should still come before AI deployment. In particular, websites and services that have been running for years without maintenance documentation should not be assumed to be secure merely because AI assisted in rewriting their interfaces or code.

For professionals: When using AI coding tools, the focus should be not only on generation speed but also on code review, dependency checks, and test coverage. Being able to hand a mature solution to an agent does not mean you can determine whether the result is fit for production.

For the consumer market: Website hijacking, supply-chain compromise, and service outages still fall on everyday users. AI tools allow digital products to launch faster, but they also allow security failures to spread faster. Defensive capabilities must keep pace with deployment speed.