Summary Points
- SGLang flaw enabled a breach: An attacker ran code on a NEAR AI inference server and may have viewed requests processed there—even in confidential mode.
- Some data was accessed: About 100 users’ account details and data from 18 older AI agents were taken; two customers’ API keys were exposed and revoked.
- No evidence of Private Chat content theft: NEAR AI found no sign that stored conversations or uploaded files were accessed.
- NEAR AI secured systems and restored service: It patched servers, rotated credentials, and notified affected users; anyone not contacted need take no action, but should watch for suspicious account-related messages.
Code flaw exposed AI systems
An attacker used a flaw in SGLang, an open-source tool that helps run AI models, to run code on a NEAR AI server. The breach showed that attackers can exploit software inside a secure environment, even when that environment aims to keep data private.
NEAR AI said the attacker could have viewed requests handled by the server, including requests sent in confidential mode. The company has not said whether those requests included chats handled for Brave’s Leo assistant.
Data accessed as company responds
The attacker also reached an older NEAR AI system and downloaded account details for about 100 users. The information included names and email addresses, and, for some users, plan and balance details. NEAR AI found no evidence that the attacker accessed stored Private Chat conversations or uploaded files.
The attacker also took chat histories, settings and logs from 18 AI agents. Logs exposed two customers’ cloud API keys, which NEAR AI has revoked. The company also shut down Private Chat, reset credentials and agent tokens, fixed the SGLang flaw on its servers, and restored its cloud service.
The incident highlights a practical limit of confidential computing: it cannot guard against every weakness in the software running inside a protected system. As AI assistants take on more personal tasks, strong security must protect both the environment and the programs within it. NEAR AI says affected users have received direct notice; others do not need to act.
Discover More Technology Insights
Stay informed on the revolutionary breakthroughs in Quantum Computing research.
Explore past and present digital transformations on the Internet Archive.
DataProtection-V1
