Deep Dive into the NEAR AI Server Security Incident

NEAR AI has released a detailed report concerning a security breach that impacted its infrastructure, stemming from a vulnerability in a third-party open-source component. The incident centered on a remote code execution flaw within the SGLang inference engine running on their servers.

The Attack Vector and Scope of Data Compromise

The attack was executed on November 8th. By exploiting the SGLang vulnerability, the threat actor achieved code execution on a推理 server and gained a foothold in internal systems. The subsequent investigation confirmed the following data was accessed:

  • User Account Information: Approximately 100 users' personal data, including email addresses and names, were downloaded from a deprecated Private Chat backend. Some records also contained subscription plan details and account balances.
  • Agent Data Exposure: The attacker accessed 18 older版本的 agents that were not running within a TEE (Trusted Execution Environment) secured enclosure. Their chat histories, configuration settings, and operational logs were obtained.
  • API Key Leakage: API keys belonging to two Cloud customers were exposed as they were present in log files. These keys have been revoked.

The report also noted that the attacker potentially could observe user requests being processed by the compromised server during the intrusion period, even those in a confidential mode. However, NEAR AI stated there was no evidence that stored Private Chat conversation histories or user-uploaded files were accessed.

Root Cause and Security Limitations

The primary cause was the vulnerability in the SGLang engine. NEAR AI clarified a crucial distinction in their report: while TEE technology is designed to protect data from the server operator or host provider, it does not safeguard against vulnerabilities within the application software running inside the TEE itself. An application-layer flaw remains exploitable even within a TEE.

Incident Response and Remediation Actions

Upon detecting the breach, NEAR AI's security team activated emergency protocols, implementing several decisive actions to contain the incident and strengthen defenses:

  • Immediately revoked all active user sessions and administrative credentials.
  • Issued new security tokens for all AI agents.
  • Permanently decommissioned the at-risk Private Chat service.
  • Deployed patches for the SGLang vulnerability across all inference servers.
  • Conducted a full credential rotation and forensic review.

NEAR AI has notified all confirmed affected users and reported the vulnerability details to the SGLang open-source project to aid broader community security.