near ai breach: SGLang Bug Exposed 100 Users and 18 Agents at a Private-AI Company

The near ai breach is a rough look for a company selling private AI: NEAR AI says an attacker used a remote code execution bug in SGLang, the open-source inference engine running its models, to break into an inference server and move deeper into its systems. Names and emails for about 100 users, data from 18 AI agents and two customers' API keys were exposed, according to its November 9 incident report.

What actually happened

NEAR AI, part of the wider NEAR ecosystem, runs AI products including a Cloud service that uses trusted execution environments, or TEEs, to keep prompts encrypted and isolated from server operators. In its report, summarized by Decrypt and Cointelegraph, the company said an attacker exploited a remote code execution vulnerability in SGLang to run commands on one inference server, then reached other parts of the infrastructure.

Investigators traced the intrusion to an internal system linked to Private Chat, a feature NEAR AI had already retired. From there the attacker downloaded account records for roughly 100 users. Separately, 18 agents on an older deployment of NEAR AI Agents were accessed, including their chat histories, settings and logs. Those agents were not running inside the confidential computing environment.

Logs also held NEAR AI Cloud API keys belonging to two customers. NEAR AI says both keys were revoked but did not say whether they were used.

"On detection, we revoked all active sessions and administrative credentials, replaced all agent tokens, took Private Chat offline, and deployed a fix for the SGLang vulnerability across all inference servers," the team said, per Decrypt. It has notified affected users and contacted all 18 agent owners individually.

Why the near ai breach matters

The big lesson is in one line from NEAR AI's own report: hardware isolation "does not protect against a vulnerability in the software running inside the environment." A TEE stops the host from peeking, but if the inference engine inside it is exploitable, an attacker can run code right where the sensitive requests live.

That is why the possible exposure of November 8 confidential-mode requests stings. It is a possibility the investigation flagged, not a confirmed theft, but it goes to the heart of the "private AI" pitch.

SGLang itself was already under scrutiny. Decrypt notes that Carnegie Mellon's CERT Coordination Center documented six SGLang vulnerabilities in July. NEAR AI has not named the specific flaw used here and reported it to SGLang's developers, so there is no basis to tie this breach to those earlier bugs.

There is a partner angle too. In October, NEAR AI said Brave's Leo assistant used its confidential inference for a supported model. The incident report does not identify Brave users among those affected.

Not the only NEAR headline this week

Different story, same ecosystem: NEAR also came up in our near laundering memecoins piece about throwaway tokens used to move hack funds. The two incidents are unrelated. On the AI security side, labs are racing to harden their stacks, as we saw in cathedral systems openai.

What we still don't know

NEAR AI has not said how long the attacker had access, which SGLang version was exploited, or whether the two exposed API keys were used before revocation. We did not review the original incident report ourselves, so the details above come from NEAR AI's statements as summarized by Decrypt and Cointelegraph.

What to do as a reader

If you have a NEAR AI account, watch your inbox for an official notice and be wary of unsolicited messages about your balance or account, which NEAR AI itself warned about. Rotate any API keys you used with NEAR AI Cloud and check the NEAR AI account on X for updates.

If you run SGLang yourself, update to the latest version, lock down network exposure and review the July CERT advisories. The near ai breach shows that "confidential" does not mean "invulnerable." Nothing here is investment advice.