SIAgentNews.com
← All news
DevelopingAgentsSecurity

NEAR AI data breach exposes 100 users via SGLang flaw

Attackers exploited a remote code execution vulnerability in inference software to access user emails, agent logs, and API keys.

· First reported by CryptoTicker.io

100approximate number of users affected

The short version

  • Attackers exploited a remote code execution vulnerability in SGLang, the open-source software used to run AI models.
  • Data exposed includes names, emails, account balances, and chat histories for 18 AI agents.
  • Two cloud customer API keys were also accessed, though NEAR AI has since revoked them.
  • NEAR Protocol tokens, wallets, and private keys remain unaffected by this service-level breach.

Why it matters

This incident highlights that even 'confidential computing' environments can be compromised if the software running inside them has flaws. For users, the primary risk is now sophisticated phishing using the leaked personal details.

The SGLang vulnerability

An attacker recently exploited a remote code execution vulnerability within SGLang, an open-source inference engine. Inference is the specific step where a trained AI model processes a user request to produce an answer. Because the engine must see the text to process it, any attacker who can run code on the server can see the data in plain text.

NEAR AI uses SGLang to run its language models on its own servers. The intruder managed to execute commands on a single inference server and then moved deeper into the company's infrastructure. This allowed them to reach internal systems, including a discontinued feature called Private Chat and an older deployment of AI agents.

The company has since patched the flaw on all inference servers and revoked all active sessions. They also replaced every agent token and rotated all remaining credentials. While NEAR AI has not named the specific SGLang vulnerability, they have reported the issue to the software's developers for a fix.

What data was taken

The breach resulted in the confirmed theft of account data for around 100 users. This information includes names and email addresses, and in some instances, the users' subscribed plans and their account balances. It is important to note that these balances refer to service credits for AI calls, not actual cryptocurrency in a wallet.

The attacker also accessed the chat histories, settings, and logs for 18 AI agents on an older platform. Additionally, the logs contained API keys for two cloud customers. An API key is a digital access code that allows programs to log into a service, and NEAR AI has already revoked both of these keys.

There is a residual risk regarding requests processed on November 8. The attacker may have observed requests being processed on the compromised server during that window, even those in confidential mode. However, NEAR AI reported finding no evidence that they accessed stored files or conversations from the Private Chat feature.

The limits of confidential computing

NEAR AI's business model relies on confidential computing, which uses a trusted execution environment (TEE). A TEE is a hardware-isolated area of a processor designed to keep data encrypted and hidden even from the server operator. This is intended to ensure that prompts stay private while the model is processing them.

This incident demonstrates a critical limitation of this technology. While a TEE protects data from the host and the operator, it does not protect against vulnerabilities in the software running inside the environment. The SGLang flaw acted like a trusted employee letting a stranger into a secure vault, rendering the hardware isolation ineffective.

Moving forward, NEAR AI plans to audit its inference software and surrounding services as strictly as the hardware environments themselves. The company maintains that its code and configurations remain public so that third parties can verify what is running within the isolated environment. This move aims to restore trust in their privacy-focused AI computing power.

Risks and next steps

The most immediate danger for affected users is targeted phishing. Because scammers now have names, emails, and account balances, they can craft highly convincing messages. These messages might pose as official security checks, offer fake compensation in tokens, or request a 'migration' to a new address to move balances.

Users are advised to treat any unsolicited messages regarding their account with extreme caution. NEAR AI has contacted the affected users and the owners of the 18 agents directly. The company warns that no reputable provider will ever ask for a seed phrase or private key through email, phone, or chat.

For those using NEAR AI Cloud, it is recommended to create new API keys and delete the old ones immediately. Users should also renew their passwords and enable two-factor authentication where available. If a user has used their password on other services, they should change it there as well to prevent further unauthorized access.

What to watch

  • Whether the attacker successfully used the two stolen API keys.
  • How many users fall victim to phishing scams using the leaked account data.
  • The results of the upcoming audits on NEAR AI's inference software.

Try it

If you use NEAR AI Cloud, manually generate and replace your API keys today, even if you weren't told your account was specifically compromised.

Sources and links

First reported by CryptoTicker.io. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news