The infrastructure of the NEAR AI project suffered a targeted cyberattack, as a result of which an attacker gained unauthorized access to internal systems. The vector of penetration was a vulnerability in the SGLang software — a framework used for serving large language models and deploying AI agents.

According to confirmed data from the project team, the scope of the compromise affected the information of approximately 100 users, as well as the data of 18 AI agents operating in the NEAR AI ecosystem. The incident demonstrates the growing attack surface for blockchain projects that integrate generative artificial intelligence into their infrastructure.

Technical nature of the threat

SGLang is a high-performance framework for inference and serving of language models, widely used in decentralized AI solutions. A vulnerability in such a layer creates a critical risk: an attacker who gains control over an inference server is potentially able to intercept requests, manipulate agent responses, and gain access to sensitive data of interacting users.

Such incidents highlight a fundamental problem at the intersection of Web3 and AI — the security of infrastructure software becomes no less important than the security of smart contracts. The attack on NEAR AI illustrates that even projects with a strong blockchain architecture remain vulnerable through third-party dependencies in the AI tooling stack.

Cryptalist analysis: The incident at NEAR AI signals the emergence of a new class of risks for crypto ecosystems — compromise through AI infrastructure rather than through classic blockchain vulnerabilities. The market should expect increased attention from regulators and investors to the security of decentralized AI services, which could lead to a reassessment of risks in the AI token segment and stricter requirements for auditing infrastructure software.