NEAR AI brought its cloud inference service back online on November 10, hours after detecting unauthorized access to internal systems and suspending operations as a precaution. According to the company, it closed the unauthorized access, patched vulnerabilities across its servers, and replaced affected credentials before restoring the platform.

The service resumed approximately four hours after NEAR AI's initial public disclosure. The company confirmed that NEAR Protocol and other NEAR services were unaffected by the incident.

The breach was first disclosed late on November 9, when NEAR AI suspended cloud inference while investigating the unauthorized access. A subsequent security update confirmed that the intrusion had been closed, although cloud operations remained suspended until the infrastructure could be safely restored.

NEAR AI has not disclosed how the attacker gained access, which internal systems were compromised, or how long the intrusion lasted.

Customer Prompts Protected by Hardware Isolation

NEAR AI emphasized that its confidential inference workloads operate inside hardware-isolated trusted execution environments (TEEs), which are designed to prevent unauthorized access to customer prompts during processing. The company's private inference infrastructure uses Intel TDX (Trust Domain Extensions) and NVIDIA confidential computing to isolate sensitive workloads from the underlying host environment. Confidential computing describes this category of hardware-based protection, which shields data while it is actively being processed and complements conventional encryption that protects data at rest and in transit.

Prompts are encrypted and processed inside protected enclaves, preventing infrastructure operators — and NEAR AI itself — from reading their contents under normal operating conditions. Cryptographic attestation additionally allows users to verify that supported inference requests were processed within the intended secure environment.

These protections separate access to internal administrative systems from access to customer prompts processed inside secure enclaves. Whether the intrusion stopped at that administrative boundary is the central open question of the incident. Even so, NEAR AI's initial announcements did not establish the full scope of the intrusion or provide a comprehensive assessment of potentially exposed information. The company's service terms acknowledge that hardware isolation reduces security risks without eliminating vulnerabilities associated with the underlying hardware, firmware, or execution environment.

NEAR AI Cloud provides confidential AI inference to developers and applications, including Brave's Leo AI assistant, which expanded its use of NEAR AI's protected infrastructure in October. The service also supports private AI models used for cybersecurity research, following the company's October 29 launch of additional security-focused models and its NEAR AI Red initiative. That customer base is what gives the incident relevance beyond NEAR AI itself: the platform's protected inference environment is an operational dependency for the applications built on it, and it was offline for several hours.

The interruption was confined to NEAR AI's cloud infrastructure. The company reported no disruption to the NEAR Protocol blockchain, which operates independently of its hosted inference services.

NEAR AI has not identified the party responsible for the intrusion or disclosed whether additional customer-facing security measures will be required. Its latest announcement confirmed that the affected vulnerabilities had been fixed on every server and that a full incident report would be released shortly. The promised report is the next development to watch, along with any company statement on additional customer-facing security measures.