highresolved

NEAR AI inference-server compromise exposes account and agent data

NEAR AI reported an SGLang execution flaw that enabled access to an inference server and internal services. The disclosed impact included account details, legacy agent histories and two customers’ API keys.

NEAR AI said the attacker reached a retired chat backend and 18 agents on a legacy deployment. Requests being processed on the accessed inference server on November 8 could also have been observed, including confidential-mode requests.

The November 9 report says access was closed, the flaw fixed and credentials replaced. The company distinguished data exposed during processing from stored chat conversations and uploaded files, for which it found no indication of access.

Takeaway

Patch the inference engine as well as the surrounding infrastructure. Hardware isolation does not compensate for an exploitable service inside the protected environment.

Sources

ainear-ai