Zero-Click AI Attacks as Emerging Cyber Threat
Saalbach, Klaus · osnaDocs (Osnabrück University) · 2025
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
osnaDocs (Osnabrück University) · Osnabrück University · report · 2025
Abstract
This working paper analyzes Zero-Click AI Attacks as a new and emerging cyber threat, shows the characteristics of this attack methods and its variants as well as the cyber security measures. Artificial Intelligence (AI) is commonly understood as the ability of machines to perform tasks that normally require human intelligence. A widespread AI application is the Generative AI (GenAI) with Large Language Models (LLMs) that create content based on short instructions (prompts), which are a key vulnerability if malicious instructions are given.
Attempts to circumvent restrictions (jailbreaks) are done by prompt injections, i.e., special instructions to AI to provide restricted or malicious content. Typical prompt injections are direct commands like DAN (do anything now), misleading the AI by fictional situations, and reverse psychology to facilitate the creation of malware, viruses, ransomware, and other malicious applications. AI Agents are software programs where LLMs with reasoning ability have increasing autonomy to achieve given objectives.
An important feature of AI Agents is the retrieval augmented generation (RAG), where LLMs can access external sources like documents, web sources, or external databases to generate information. Originally, cyber attacks were based on computers that were used to attack other computers. Then, attackers used prompt injections to use AI applications against their targets. In 2025, a new cyber threat emerged where the attacker does not use its own AI against another computer, but attacks instead the local AI model on the target computer.
This new attack method is characterized by zero clicks, i.e., the victim does not need to click on anything to execute the attack, and by hidden text which is not legible for humans, but legible for the AI (like white-on-white letters with tiny font of 1pt size). The hidden text (word text or HTML) can be embedded into emails, PDF files, images, videos, websites etc. and contains prompts for the local AI model on the victim computer to exfiltrate sensitive data and to execute other malicious commands. As these prompts are not directly given by the attacker to the targeted AI, they are known as indirect prompt injections.
The most prominent examples are EchoLeak and AgentFlayer which exploit the connection between AI Agents and other applications to get access to these applications and their data. However, evaluations in August 2025 have shown that there are lots of possible attack variants and potential targets. This is a relevant cyber threat as every computer with a local AI application/AI Agent can be approached while conventional cyber security is circumvented. Another aspect is the simplicity of attacks and the limited efforts needed by the attackers.
Cyber security countermeasures include detection and removal of hidden elements and instructions (prompt sanitization), definition of authorized users who are allowed to prompt the AI, restricting and monitoring of AI connector functions, blocking of AI-generated emails, repositories of malicious websites or whitelisting with use of safe websites only and lists of malicious prompts. In summary, the rapid expansion of AI, AI Agents, multi-agentic systems, and their capabilities results in a rapid growth of cyber threats and attack variants which is an important cyber security issue.
Open the resolved locator · 10.48693/759
Metadata licence: OpenAlex metadata (CC0)