How AI Is Reshaping Phishing: What German Companies Face Now
Phishing has always been a game of deception. Attackers impersonate trusted entities to trick people into handing over credentials, money, or access. But artificial intelligence is making that deception far more convincing—and far harder to detect.
Recent data from Germany paints a stark picture. According to a report cited by Hardbloxx, 87 percent of German companies have been affected by AI-based phishing. Total phishing attempts jumped 150 percent year over year. Voice phishing, or vishing, now succeeds 40 percent more often than traditional email-based attacks because AI can clone a person’s voice from just a few seconds of audio.
AI is not only generating more attacks but better ones. ThreatLabz, a security research team, identified over 413,000 AI-generated phishing websites. Attackers use text-to-site tools to build professional-looking portals in minutes. The same report found that 95.2 percent of phishing attempts now hide in encrypted traffic, making them invisible to traditional inspection tools.
The Shift from Spray-and-Pray to Precision
Older phishing campaigns cast a wide net. A generic email about a compromised account would go to millions of recipients. AI changes that. Large language models let attackers craft personalized messages that reference a target’s job, recent purchases, or internal company language. That personalization drives higher click rates and fewer obvious red flags.
Session hijacking has also evolved. Even when users have multi-factor authentication enabled, attackers can steal session cookies after the user logs in. This bypasses MFA entirely. ThreatLabz notes that session hijacking techniques are now common among AI-assisted phishing operations.
The Botnet and IoT Connection
AI is not limited to email and websites. Botnets are becoming smarter. One recent example is Dysphoria, a botnet that compromised nearly 296,000 IoT devices. Those devices were used for distributed denial-of-service attacks and as residential proxies, helping attackers hide their true locations. Another botnet, ToxNetV2, integrates a large language model into its decision-making cycle, allowing it to adapt its behavior on the fly.
Detection Research: Hybrid Models
Researchers are working on countermeasures. A comprehensive review published on Springer Professional examines hybrid detection models that combine machine learning with rule-based systems. These models analyze website features, URL structure, and content patterns to flag phishing sites before they go live. The approach aims to catch AI-generated pages that look legitimate to the naked eye but contain subtle anomalies in underlying code or behavior.
What Companies Must Do Now
The German Federal Office for Information Security (BSI) has issued guidance that treats AI agents as privileged users. That means applying the same controls: multi-factor authentication, strict access logging, and regular activity audits. A survey cited in the Hardbloxx report found that 68 percent of companies already use AI agents with extensive access rights, and 55 percent have experienced security incidents involving those agents.
Companies should also assume that encrypted traffic can still carry threats. Modern security tools must inspect SSL/TLS traffic for phishing payloads, even if that adds latency. Employee training needs updating too. Classic phishing indicators—typos, generic greetings, bad grammar—no longer work as reliable signals. AI generates fluent, error-free text.
The German government is establishing a national AI safety institute to coordinate research and policy. But for individual organizations, the clock is already ticking. AI-powered phishing will only grow more precise and more invisible. The question is whether defenses can keep pace.
This wave of phishing is not about tricking humans into spotting a misspelled URL. It is about building systems that can detect AI-generated deception at scale. That requires new detection models, stricter access controls, and a sober acceptance that the old rules no longer apply.