NEAR AI says an attacker exploited a remote code execution vulnerability in SGLang, the open-source inference engine running on its servers, to gain access to part of the company’s infrastructure. The attacker first obtained code execution on one inference server, then reached an internal system behind the company’s previously discontinued Private Chat service, according to NEAR AI’s security incident report, dated November 9.
The company said the attacker downloaded account details for approximately 100 users, including email addresses and names, with plan and balance information included for some accounts. The intrusion also reached 18 agents on a legacy NEAR AI Agents deployment that was not running inside the company’s confidential-computing environment. Data retrieved from those agents included chat histories, settings and logs. The account-data exposure differs from the public cloud-storage misconfiguration involved in the 302.ai data leak.
Two NEAR AI Cloud customers also had API keys exposed in logs, and the company says both keys have been revoked. The incident highlights the risk of credentials appearing in operational records, a problem also seen in the stolen-token activity described in LLMjacking attacks against cloud-hosted AI services.
NEAR AI said the attacker could have observed user requests being processed on the compromised server on November 8, including requests running in confidential mode. The company found no indication that stored Private Chat conversations or uploaded files were accessed. It said the 18 legacy agents were separately accessed and their chat histories, settings and logs were retrieved.
After detecting the intrusion, NEAR AI revoked active sessions and administrative credentials, replaced agent tokens, took Private Chat offline and deployed a fix for the SGLang vulnerability across all inference servers. It said it completed a forensic review, rotated the remaining credentials, notified users whose accounts or agents were confirmed to be affected and restored access to NEAR AI Cloud. The company has also reported the vulnerability to the SGLang project.
NEAR AI said confidential computing protects data from the operator and host, but not from a vulnerability in software running inside the protected environment. In this incident, the SGLang flaw allowed the attacker to execute code on an inference server and access requests while they were being processed, including requests handled in confidential mode.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
