Category: AI Agent Security | Reading time: 3 min
NEAR AI published an incident report on November 9 saying an attacker exploited a remote code execution flaw in SGLang, the open-source engine it uses to serve AI models, to run code on one of its inference servers and move into its infrastructure. The attacker retrieved the chat histories, settings and logs of 18 customer AI agents.
What was accessed
According to NEAR AI's report:
- 18 agents on a legacy NEAR AI Agents deployment that did not run inside a trusted execution environment (TEE, a hardware-isolated enclave). Their chat history, settings and logs were retrieved. - About 100 users' account details, including email, name and some plan or balance information, downloaded through an internal system behind its deprecated Private Chat product. - Two NEAR AI Cloud customers' API keys, which were exposed in logs and have been revoked. - On November 8, the attacker could have observed requests processed on the affected server, including requests sent in confidential mode.
NEAR AI says it has no indication that stored Private Chat conversations or files were accessed. The report does not name the attacker or the SGLang CVE.
The response
NEAR AI says it revoked sessions and admin credentials, replaced agent tokens, took Private Chat offline, deployed an SGLang fix, reported the flaw to the SGLang project, started a full credential rotation and forensic review, and is making regulatory notifications where required.
Why it matters for agent deployers
An agent's chat history, settings and logs are a record of what it was told, what it could reach and what it did. In this case that record was the thing taken. NEAR AI's own takeaway is that an isolated enclave does not protect data if the software running inside it has a hole.
For teams running agents on third-party model infrastructure, the incident is a reminder that the serving stack is part of the attack surface. Credentials and tokens visible to agents and their logs should be short-lived and easy to rotate. Teams also need their own record of what each agent did, kept outside the provider's systems, so they can tell customers and regulators exactly what was exposed.
This account comes from NEAR AI's own report and has not been independently confirmed.
Source: NEAR AI: Security incident report, SGLang vulnerability