NEAR AI has disclosed that it detected and shut down unauthorized access to portions of its systems following an intrusion on November 8, 2026. An attacker leveraged a remote code execution flaw in SGLang, the open-source software used to run artificial intelligence models on the company’s inference servers.
This allowed the intruder to execute code on a single inference server and then move into connected internal infrastructure.
Investigators determined that the attacker reached systems linked to a previously retired Private Chat feature and obtained account records for roughly 100 users.
The downloaded information consisted of email addresses and names, along with subscription plan and balance details in certain cases.
Separately, the intrusion reached 18 agents operating on an older deployment of NEAR AI Agents.
Those agents were not protected by the company’s confidential computing setup, which relies on trusted execution environments (TEEs).
Data taken from them included conversation histories, configuration settings, and operational logs. Logs also contained API keys belonging to two NEAR AI Cloud customers; those keys have since been revoked.
On the day of the intrusion, the attacker may have been able to observe user requests as they were processed on the compromised server.
This potentially included requests handled in confidential mode.
NEAR AI reported no evidence that previously stored Private Chat conversations or uploaded files were accessed.
The company emphasized a core limitation of its security model: confidential computing shields data from the platform operator and the underlying host infrastructure, yet it cannot defend against vulnerabilities in the software running inside the protected environment.
Because the SGLang flaw allowed code execution within one inference server, requests in transit on that server—including confidential ones—could be exposed.
In response, NEAR AI revoked active sessions and administrative credentials, replaced agent tokens, took the Private Chat service offline, and applied a fix for the SGLang vulnerability across all inference servers.
It completed credential rotations, conducted a forensic review, and directly notified affected users. Access to NEAR AI Cloud has been restored.
The firm reported the vulnerability to the SGLang project and stated it is now applying the same security scrutiny to inference software and surrounding services as it does to the enclaves themselves.
The underlying NEAR Protocol blockchain was not affected.This incident underscores persistent challenges in AI infrastructure security.
Open-source inference engines power much of the current ecosystem, yet they introduce supply-chain and software-layer risks that hardware isolation alone cannot fully mitigate. Similar pressures have appeared elsewhere.
In a separate October 2026 incident, attackers compromised infrastructure at the cloud development platform Railway, targeting AI provider credentials.
The intrusion potentially affected environment variables for up to approximately 100,000 workspaces and led to roughly $740,000 in unauthorized AI usage charges.
Investigators linked the activity to a broader campaign against platforms that store AI keys, sometimes referred to as “Black October,” in which stolen credentials were used for model distillation or resale.
Other recent issues have highlighted software weaknesses inside AI serving stacks, including sandbox-escape flaws in language runtimes that could affect engines such as vLLM.
Legacy deployments, credentials appearing in operational logs, and the rapid integration of agentic systems further expand the attack surface.
As organizations adopt confidential AI offerings that promise privacy even from the provider, the NEAR AI case illustrates that TEEs address only part of the threat model.
Rigorous patching of inference software, minimization of legacy components, careful secret management, and continuous verification of the full software stack remain essential. Users who have not been contacted by NEAR AI are advised that no action is required, though caution toward unsolicited messages referencing account details is prudent.
