A vulnerability discovered in SGLang, an open-source language model serving framework, has exposed critical infrastructure within NEAR's ecosystem, compromising data from 18 distinct AI agents. The breach demonstrates how even sophisticated systems built with hardware-backed security assumptions can fail when application-layer vulnerabilities go unpatched. An attacker leveraged the flaw to gain unauthorized access to internal systems, potentially intercepting confidential requests flowing through the AI pipeline—a particularly sensitive exposure given the proliferation of sensitive data now coursing through decentralized AI networks.
The incident underscores a troubling pattern in blockchain infrastructure security: architectural layers that appear isolated often share attack surfaces that remain invisible until exploitation. Hardware-backed attestation and secure enclaves provide valuable cryptographic guarantees, but they operate within a broader system where software flaws can render those protections moot. In this case, the SGLang vulnerability allowed an attacker to circumvent application-level access controls entirely, gaining visibility into the confidential prompts and model interactions that users believed were protected by NEAR's infrastructure. This type of breach is particularly damaging because it violates user expectations about privacy; many builders integrating AI into blockchain applications assume that deploying on established chains provides baseline security guarantees that, as this incident shows, may not materialize in practice.
The exposure of 18 AI agents suggests this was not an isolated misconfiguration but rather a systematic problem affecting multiple applications built on top of the compromised infrastructure. This raises urgent questions about dependency management and security auditing in the Web3 AI stack. As AI becomes increasingly central to DeFi, oracle networks, and autonomous agent ecosystems, the security posture of foundational components takes on existential importance. A single unpatched framework can create cascading failures across dozens of downstream applications, each assuming their upstream dependencies have been sufficiently hardened.
NEAR's response and the broader industry's reckoning with this vulnerability will likely accelerate conversations around formal verification, staged rollouts, and mandatory security audits for AI infrastructure components. The breach illustrates that hardware security primitives and cryptographic design are necessary but not sufficient; the entire stack requires defense-in-depth, from kernel-level hardening through application-level access controls and continuous monitoring.