Near AI Infrastructure Compromised via SGLang Inference Engine Vulnerability
A remote code execution flaw exposed user data and legacy AI agent configurations.
1Security Breach Overview
Near AI confirmed a security breach occurring on November 8, 2026, involving a remote code execution vulnerability in the SGLang inference engine. An unauthorized party exploited this flaw to gain access to internal infrastructure. The incident resulted in the exposure of account details for approximately 100 users and 18 legacy AI agents.
2Data Exposure Details
The intruder accessed a deprecated Private Chat service, downloading names, email addresses, and subscription balance information for 100 users. Additionally, 18 legacy AI agents operating outside trusted execution environments had their chat histories, configuration settings, and logs compromised. Two Near AI Cloud customers also suffered the exposure of their API keys.
3Confidential Computing Limitations
The breach highlights a critical limitation in confidential computing technology. While hardware-isolated execution environments protect data from hosts, they cannot defend against software-level vulnerabilities within the applications themselves. Because the SGLang engine ran inside the trusted environment, the attacker successfully bypassed hardware isolation to observe confidential AI requests during active processing.
4Immediate Remediation Actions
Upon detecting the intrusion, Near AI revoked all active sessions and administrative credentials. The company replaced compromised agent tokens, took the Private Chat service offline, and deployed a patch for the SGLang vulnerability across all inference servers. Furthermore, engineers completed a full forensic review and rotated all remaining system credentials to ensure security.
5Notification and Communication
Near AI has individually contacted the owners of the 18 compromised agents and all affected users via [email protected]. The company advised users to remain vigilant against unsolicited communications regarding account balances or personal information. Users who have not received direct notification from the company are currently considered unaffected by this specific security incident.
6Scope and Ecosystem Impact
The company clarified that the breach did not affect the Near Protocol itself or other broader ecosystem services. While the incident involved Near AI Cloud infrastructure, the company has not confirmed whether the compromised server handled chats for external partners like Brave’s Leo assistant. Forensic analysis remains focused on the specific inference software vulnerabilities.
7Technical Disclosure Status
Near AI reported the SGLang flaw to the project maintainers to facilitate a broader fix. However, the company has not yet identified the specific vulnerability by a CVE number or named the affected software versions. This lack of granular technical detail leaves questions regarding the full extent of the potential exposure for other users.
8Future Security Implications
This incident forces a re-evaluation of how AI infrastructure providers secure their processing layers. Near AI stated that future security audits will apply the same rigor to inference software as they do to hardware-protected environments. The breach underscores the necessity of securing the entire software stack, not just the underlying isolated execution hardware.