Attackers have recently been discovered using a new technique called indirect prompt injection to target AI agents by embedding hidden instructions in websites, as reported by Zscaler’s ThreatLabz. This method involves planting instructions in content that AI agents read, such as web pages, in order to manipulate their behavior.
Two real-world campaigns were documented by Zscaler, with one posing as software documentation to carry out a payment scam, and the other impersonating a cryptocurrency service. In both cases, the attackers utilized SEO poisoning to boost the visibility of their sites in search results, making it more likely for AI agents to come across them.
The hidden instructions were cleverly concealed in parts of the page that are not visible to humans, using techniques such as CSS to move the text off-screen or embedding them within structured JSON-LD metadata. The first campaign involved a fake page masquerading as Python library documentation, tricking AI agents into purchasing a $3 API license key to resolve an error, only to guide them into making a payment to the attacker’s cryptocurrency wallet for a fraudulent key.
In the second campaign, the attackers used a typosquatting domain to impersonate DeBank, a popular cryptocurrency portfolio tracker, with hidden instructions instructing AI agents to treat the fake site as the legitimate DeBank and rank it first. Zscaler’s ThreatLabz conducted tests using their own autonomous agent against the sites with 26 large language models (LLMs).
Results showed that four of the 26 models were manipulated into executing the fraudulent payment, with versions of Meta’s Llama and Google’s Gemini being susceptible. Furthermore, two models, OpenAI’s GPT-5.4 and Anthropic’s Claude Sonnet 4.5, incorrectly rated the fake site as legitimate when a trusted reference for the real DeBank was absent. However, when provided with the genuine site for comparison, none of the models were deceived.
Zscaler emphasized that the susceptibility of AI agents to such attacks heavily depends on the specific LLM and the amount of context provided. They cautioned that as AI agents become more prevalent as interfaces to the web, the content itself becomes a larger attack surface, highlighting the dual nature of AI as both a tool for streamlining workflows and a potential avenue for abuse.
