An attacker exploited a remote-code-execution vulnerability in SGLang, the open-source inference engine running on NEAR AI's servers, exposing about 100 accounts and data from 18 agents. The company's own incident report concedes its confidential-compute enclave did not stop it. That admission rewrites the checklist every builder uses when buying "private inference."
Key facts
- The intrusion came through a remote-code-execution vulnerability in SGLang, the open-source inference engine, giving the attacker code execution on one inference server and a path deeper into NEAR AI's infrastructure.
- Account details (email addresses, names, and for some accounts plan and balance information) were downloaded for approximately 100 users.
- Data from 18 agents on a legacy non-TEE deployment was retrieved, including chat history, settings, and logs.
- Two Cloud customers' API keys were exposed in logs and have been revoked.
- On November 8, requests being processed on the accessed server could be observed, including requests running in confidential mode.
- No CVE identifier or affected SGLang version has been published; NEAR AI says it reported the bug to the SGLang project.
What happened
NEAR AI published its incident report on November 9, 2026, five days after the intrusion it dates to November 8. The sequence it describes: an unauthorized party exploited a remote-code-execution vulnerability in SGLang to run code on one inference server, then used that foothold to reach an internal system sitting behind the company's previously deprecated Private Chat feature. From there they downloaded account details for approximately 100 users. Separately, the attackers reached 18 agents on a legacy deployment of NEAR AI Agents that were not running inside the company's confidential-computing environment, and pulled chat history, settings, and logs for those agents.
The sharpest detail for anyone evaluating inference providers is not the account data. It is the server access. On November 8, according to the report, the attacker could observe user requests while they were being processed on the compromised server, including requests that were running in confidential mode. NEAR AI says it found no indication that stored Private Chat conversations or uploaded files were accessed, and that the two exposed API keys have been revoked.
NEAR AI's response, per its own account: on detection it 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. It says it has completed credential rotation and a full forensic review, has notified affected users directly from security@near.ai, and will make any required regulatory notifications.
What the report does not contain matters as much as what it does. There is no CVE identifier, no affected SGLang version range, and no statement on whether a patch is available upstream for other SGLang operators. NEAR AI says it reported the vulnerability to the SGLang project. Until an advisory lands, builders running SGLang in production have a confirmed real-world exploitation but no published indicator of which versions are safe. The Hacker News covered the incident on November 10, corroborating the scope figures from NEAR AI's own report.
Why it matters
The load-bearing sentence in the report is NEAR AI's own framing of what confidential computing does and does not do: "Confidential computing protects data from the operator and the host. It does not protect against a vulnerability in the software running inside the environment, and that is what happened here."
That is an unusually direct admission from a provider whose private-inference pitch rests on TEEs. An enclave attests what code runs on the silicon and shields memory from the host operator. It cannot vouch for that code being free of bugs. The SGLang flaw let the attacker execute code inside the enclave's own boundary and watch confidential requests stream through it. Hardware attestation verified the enclave was running precisely the vulnerable software it was configured to run.
For builders, the practical consequence is that confidential-compute claims and software-patching discipline are two separate lines on the evaluation checklist, and the second one does not substitute for the first. When a provider advertises TEE-backed inference, the follow-up questions are about the software inside the enclave: which engine, which version, what is the patch cadence, how quickly does a disclosed engine vulnerability get deployed. This incident gives concrete evidence for why those questions matter. We made the same structural point in our explainer on confidential AI inference: attestation covers the hardware boundary, and everything inside it is ordinary software with ordinary bugs.
It is also a live test for the category. NEAR AI shipped its NEAR AI Gateway macOS app on November 11, two days after this report, wiring TEE-backed private models into Claude Code, Codex, and OpenCode. The company is selling confidential inference and simultaneously publishing the clearest documentation yet of where that protection ends. Readers can weigh both facts directly.
Background
This is the second NEAR AI security event DeAI has covered. On November 10 we reported that NEAR AI had paused Cloud inference after a reported unauthorized access; at that point the attack vector was unknown. The new report supplies it: an RCE in SGLang, not a credential compromise or a misconfiguration. In September we covered NEAR AI Cloud's TEE-backed private inference launch, the product whose confidential mode was running on the compromised server.
The incident also lands in a crowded quarter for inference-stack security. In October, LiteLLM 1.107.0 shipped a credential stealer and RAT — an attack on the proxy layer rather than the engine. This month's event attacks the engine itself, inside an enclave. Together they sketch the actual attack surface of private inference: not just the provider's perimeter, but every layer of open-source software between the prompt and the GPU, whether that software runs in a TEE or not.
SGLang is one of the most widely deployed open-source inference engines, alongside vLLM, and it is the serving engine behind several hosted and self-hosted deployments in this beat. A confirmed in-the-wild RCE with no published CVE is relevant to anyone operating it, not just NEAR AI customers. Self-hosters running exposed endpoints were already the subject of our coverage of more than 152,000 internet-visible Ollama endpoints; engine-level vulnerabilities add a second way for a correctly firewalled deployment to become a liability if unpatched.
What's next
Three things to watch. First, the SGLang project's advisory: whether a CVE, an affected-version range, and a patch release appear, and how quickly other inference platforms that embed SGLang pick it up. Second, whether NEAR AI publishes a fuller technical post-mortem once its forensic review concludes — the current report describes the response but not the exploit mechanics. Third, whether other confidential-inference providers respond publicly, since the same engine-plus-enclave architecture is common across the category. If you run SGLang yourself, the immediate move is to verify your deployment tracks upstream fixes and to treat any version you cannot confirm as patched as untrusted until an advisory says otherwise.
Questions
- what happened in the NEAR AI SGLang security incident?
- Per NEAR AI's November 9 incident report, an unauthorized party exploited a remote-code-execution vulnerability in SGLang, the open-source inference engine, gained code execution on one inference server, and reached an internal system behind deprecated Private Chat. About 100 user accounts and 18 agents on a legacy non-TEE deployment were affected.
- does a TEE protect against vulnerabilities in the inference engine?
- No, not by itself. NEAR AI's report states confidential computing protects data from the operator and the host, but not against a vulnerability in the software running inside the enclave. The SGLang flaw let attackers run code inside one server and observe requests in processing, including confidential-mode ones.
- was there a CVE for the SGLang vulnerability used against NEAR AI?
- As of November 14, 2026, NEAR AI has not published a CVE identifier or an affected SGLang version range. The company says it reported the vulnerability to the SGLang project and deployed a fix across all of its inference servers.
- what data was exposed in the NEAR AI breach?
- According to NEAR AI: account details (email addresses, names, and for some accounts plan and balance information) for about 100 users; chat history, settings, and logs from 18 agents on a legacy non-TEE deployment; and two customers' API keys found in logs, since revoked. NEAR AI says it found no indication that stored Private Chat conversations or uploaded files were accessed.
Sources
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
