NEAR AI Breach: SGLang Flaw Exposed Confidential AI Requests
NEAR AI disclosed on 9 November 2026 that an attacker exploited a remote code execution flaw in SGLang, the open source inference engine on its servers. The intruder took account details for about 100 users and data from 18 legacy agents, and on 8 November could have observed requests on one server, including confidential mode requests. NEAR AI has patched every inference server.
On this page
What NEAR AI says the attacker reached
NEAR AI, which sells private inference on open-weight models, published a security incident report dated 9 November 2026. It says an unauthorized party exploited a remote code execution vulnerability in SGLang, the open source inference engine running on its servers. That gave the attacker code execution on one inference server, and from there access to parts of NEAR AI's infrastructure.
The report lists four kinds of exposure:
- Account details. The attacker reached an internal system behind Private Chat, a product NEAR AI had already deprecated, and downloaded details for approximately 100 users: email addresses, names, and for some accounts plan and balance information. NEAR AI says it has no evidence that stored Private Chat conversations or uploaded files were accessed.
- Legacy agents. The attacker accessed 18 agents on a legacy deployment of NEAR AI Agents and retrieved their chat history, settings, and logs. Those agents were not running in the company's confidential computing environment.
- API keys. Two Cloud customers had NEAR AI Cloud API keys exposed in logs. The keys have been revoked.
- Live requests. On 8 November the attacker could also have observed user requests being processed on the accessed server, including requests running in confidential mode.
The wording of that last point matters. NEAR AI says requests could have been observed. It does not say that any prompt was read, and it gives no count of the requests the server handled that day.
Why the enclave did not stop it
NEAR AI Cloud runs open-weight models inside Intel TDX virtual machines paired with Nvidia GPUs in confidential computing mode. The promise of that design is that the people who run the machines cannot look inside the workload.
The incident report is direct about where that promise ends. "Confidential computing protects data from the operator and the host," it says. "It does not protect against a vulnerability in the software running inside the environment," which is what happened here.
An inference engine has to see prompts in plain text to run them through a model. If an attacker can execute code inside that engine, the attacker sits where the prompts already are, and the hardware isolation around the engine is no longer in the way. The Cloud Security Alliance summed up the same point in the title of its analysis of the incident, "Inside the Enclave, Still Exposed".
Some details are still missing. The report does not name a CVE or say which SGLang versions are affected, a gap TechInformed also noted in its coverage. NEAR AI says it reported the vulnerability to the SGLang project and deployed a fix across all its inference servers, but it does not give a fixed release number. The report also does not say when the intrusion began or how long it lasted.
How NEAR AI responded and what users should do
On detection, NEAR AI says it revoked all active sessions and administrative credentials, replaced all agent tokens, and took Private Chat offline. It then rotated the remaining credentials, completed a forensic review, and notified the affected users and the owners of the 18 agents. Access to NEAR AI Cloud has been restored.
The company's advice to users is short: treat unsolicited messages that mention your account or balance with caution. Genuine notices about the incident came from the company's security address. The leaked names, emails, and balances are exactly what a convincing phishing message needs.
NEAR AI also says the code, dependencies, and deployment configurations for its confidential inference stack remain public so anyone can verify what is running. Verification shows which software is running. It cannot show that the software is free of bugs.
What this means if you run SGLang or rely on private inference
Three practical points follow from the report.
- If you self-host with SGLang, watch for the fix. No advisory with version numbers was available in the sources reviewed for this article. Until one appears, keep SGLang endpoints off the public internet, put authentication in front of them, and update as soon as the project ships a patched release.
- Rotate keys that might appear in logs. Two customers were exposed because API keys were written to logs. If you send secrets inside prompts or headers to any hosted model, assume they can end up in someone's log files.
- Read confidential computing claims precisely. An enclave narrows who can see your data. It does not remove the serving software from the list of things you trust. The same holds for a model you run on your own hardware, where the software stack is yours to patch, but the prompts never leave the machine in the first place.
This is the second breach at a provider of hosted open-weight inference in a month, after the Fireworks AI incident that exposed Notion AI chats. Both show that the provider serving the model sits in the path of every prompt.