Amazon discloses AI agent attacks targeting applications such as LiteLLM and Dify, with API keys as the objective
Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys
Amazon's threat intelligence team recorded, using decoy servers called MadPot, that attackers' AI agents were scanning AI applications exposed on the internet, stealing the model provider keys in them and adding them to a shared pool, involving products such as LiteLLM, One API, Dify, Langflow and RAGFlow. The team assesses with medium confidence that the agents run on the open-source penetration testing tool ARTEX, and states that it cannot verify the results the agents reported about themselves. The list of entry points in the records includes default administrator credentials, incomplete installations, the isCliRequest header authentication bypass and SSRF in the One API family, and known vulnerabilities such as Langflow CVE-2025-3248 and RAGFlow CVE-2026-28797. Amazon also assesses with medium confidence that the operators are multiple China-related groups, based on factors including Simplified Chinese prompts and China time zone settings.
Attackers use AI agents to automatically search for exposed AI applications and steal model keys; the known entry points and check items listed in the article can be used for self-checks.