AI Technologies , AI-Based Attacks , Generative AI

A New Self-Spreading, Zero-Click Gen AI Worm Has Arrived!

Researchers Created Worm That Can Exfiltrate Data, Spread Spam and Poison AI Models
A New Self-Spreading, Zero-Click Gen AI Worm Has Arrived!
Image: Shutterstock

Developers of collaboration software such as email and chat have been scrambling to incorporate emerging generative artificial intelligence technology into their products for the past year. Winners in this race hope to seamlessly link common office tools with AI assistants to transform the user experience.

See Also: OnDemand: AI Application Security Implementation Trends in EU

But security researchers have given the industry yet another reason for caution. They have created a zero-click, self-propagating worm that can steal personal data through applications that use AI-powered chatbots.

Dubbed Morris II in a nod to the devastating computer worm that took down a sizable chunk of the internet in 1988, the new self-replicating malware uses a prompt injection attack vector to trick generative AI-powered email assistant apps that incorporate chatbots such as ChatGPT and Gemini. This allows hackers to infiltrate victim emails to steal personal information, launch spam campaigns and poison AI models, according to researchers from Cornell University, Technion-Israel Institute of Technology and Intuit.

Victims do not have to click on anything to trigger the malicious activity: The worm is designed so that it automatically spreads the malware without human intervention. Hackers can use the worm to design "adversarial self-replicating prompts" that can convince a generative model to replicate input as output. For example, if the AI model receives a malicious prompt, it will replicate it in its output and spread it further to applications that interface with it.

"The study demonstrates that attackers can insert such prompts into inputs that, when processed by gen AI models, prompt the model to replicate the input as output (replication) and engage in malicious activities (payload)," the researchers said.

Morris II targets the current ecosystem, which comprises interconnected networks of generative AI-powered agents that interact with each other to continuously and automatically carry out tasks. Companies can integrate AI capabilities into their existing applications and connect them to remote or cloud-based generative AI models in a bid to create a "smart agent" that can interpret and carry out complex inputs. The output from these services carries out actions in a semi-automated manner with human approval or in a fully automated manner.

Generative AI-powered email assistants help answer emails, forward relevant emails to others or mark spam by analyzing the contents of the email based on predefined rules and previous engagement with senders. They also auto-generate replies, often also by interpreting attached images and not just email texts. The AI email assistant market industry is projected to reach nearly $2 billion by 2032, making it imperative for the industry to address these risks, the researchers said.

In their demonstration, the researchers set up an email system that could receive and send emails using generative AI. Then, they wrote an email that included a prompt that triggered the system to use retrieval-augmented generation, or the RAG method, which AI models use to retrieve trusted external data, to contaminate the receiving email assistant's database. The email system retrieved it and sent it to the AI model, where the worm used a jailbreak technique and forced it to exfiltrate sensitive data. It also replicated the input as output as designed, thus passing on the same instructions to other AI hosts linked to it and extending the cycle.

The adversarial prompt doesn't have to be text-only. Researchers demonstrated similar results by encoding the malware into an image attached in an email, forcing the AI-powered email assistant to forward the malicious image to other AI hosts (see: US, UK Cyber Agencies Spearhead Global AI Security Guidance.)

The number of systems using generative AI at this time is "minimal," but that number is expected to grow as tech companies make significant efforts to integrate gen AI capabilities into their existing products, effectively creating generative AI ecosystems. "Due to this fact, we expect the exposure to the attack will be increased significantly in the next few years," the researchers said.

In its current state, humans can detect the adversarial self-replicating prompt and the payload and prevent the worm from propagating to new hosts. But "the use of a human as a patch for a system's vulnerability is bad practice because end-users cannot be relied upon to compensate for existing vulnerabilities of systems," the researchers said, adding that this solution is "irrelevant to fully autonomous gen AI ecosystems."

For now, companies can secure their output to ensure that it does not consist of pieces similar to the input and does not yield the same inference. This mechanism can be deployed in the agent itself or the generative AI server, the researchers said. Countermeasures against jailbreaking also can prevent attackers from using known techniques to replicate the input into the output.

The researchers said they carried out the experiment in a laboratory setting only, and they submitted the findings to popular generative AI chatbot makers Google and OpenAI before releasing the study publicly.

About the Author

Rashmi Ramesh

Rashmi Ramesh

Assistant Editor, Global News Desk, ISMG

Ramesh has seven years of experience writing and editing stories on finance, enterprise and consumer technology, and diversity and inclusion. She has previously worked at formerly News Corp-owned TechCircle, business daily The Economic Times and The New Indian Express.

Around the Network

Our website uses cookies. Cookies enable us to provide the best experience possible and help us understand how visitors use our website. By browsing, you agree to our use of cookies.