NEAR AI restores cloud service after SGLang security flaw exposure

The SGLang Vulnerability

NEAR AI recently identified and patched a security flaw within SGLang, which is an open-source inference engine used on its servers. The incident began on November 8 when an attacker exploited a remote code execution vulnerability to gain unauthorized entry. This specific type of flaw allows an outside party to run their own code on a target server.

Once the attacker gained a foothold on one inference server, they were able to move through other parts of the company's infrastructure. They eventually reached an internal system linked to a Private Chat service that the company had previously discontinued. This access allowed the unauthorized party to download various records from the internal database.

The company has since completed a forensic review and deployed fixes across all of its inference servers. They also reported the vulnerability to the SGLang project maintainers to ensure the broader community can address the issue. NEAR AI stated that it has finished replacing all affected credentials and revoking active sessions.

Scope of the Data Breach

The security report confirmed that information belonging to approximately 100 users was exposed during the breach. This data included personal details such as names and email addresses, as well as subscription plans and account balances for some individuals. However, the company noted there was no evidence that uploaded files or stored Private Chat conversations were accessed.

The breach also impacted 18 legacy AI agents that were running on an older version of the NEAR AI Agents platform. Because these specific deployments were not running inside a Trusted Execution Environment (TEE), their chat histories, configuration settings, and operational logs were accessible to the attacker. Two customers also had their Cloud API keys exposed through these logs, leading the company to revoke those credentials.

In addition to the data theft, investigators determined that the attacker could have observed requests processed on the compromised server on November 8. This included requests running in confidential mode, though the exact number of observed requests remains unspecified. NEAR AI has since taken the Private Chat service offline as part of its remediation efforts.

Confidential Computing Limits

NEAR AI utilizes Trusted Execution Environments, or TEEs, to provide hardware-based isolation for confidential computing. This technology is designed to prevent infrastructure operators and host providers from accessing sensitive data while it is being processed. Under normal conditions, these environments act as a secure vault for AI inference and blockchain-based payment services.

The recent breach revealed a critical limitation of this technology: TEEs protect against external access but cannot stop vulnerabilities within the applications running inside them. Because the SGLang flaw existed within the software itself, the attacker was able to execute code inside the protected environment. This allowed them to potentially see requests as they were being processed by the software.

Moving forward, NEAR AI plans to expand its security reviews to include inference software, related dependencies, and deployment configurations. The company is doubling down on these checks as it continues to grow its confidential computing operations. This includes a service that recently surpassed one million NEAR staked for funding AI inference.

Response and Next Steps

Following the discovery of the unauthorized access, NEAR AI initially suspended its Cloud inference operations. The company kept services paused while engineers worked to close the attacker's access and perform a full forensic review. Once the security work was completed and credentials were rotated, the Cloud infrastructure was returned to online status.

The company has contacted the specific users and agent owners whose accounts were confirmed to be accessed. NEAR AI told affected users, "If you haven’t heard from us, no action is needed," in a public update. They advised all users to remain cautious regarding any unsolicited messages that reference personal account information or balances.

While this incident occurred during a week involving separate reports of NEAR-based tokens being used in a Binance-related hack, NEAR AI confirmed no connection exists between the two. The company stated that its underlying blockchain protocol remained unaffected throughout the SGLang incident. Regulatory notifications will follow for the company where they are required by law.