NEAR AI was breached through the engine running its agents, not the prompt

NEAR AI, an AI platform, suffered a breach due to a flaw in its model-serving engine, SGLang, not from the prompt itself. The attacker obtained code execution on one of NEAR AI's inference servers, then navigated deeper into the infrastructure using an internal system connected to a retired feature. The breach exposed personal information of approximately 100 users, payment details for some users, chat histories, settings, and logs from 18 older agents, as well as AWS API keys for two customers.

These compromised keys have since been revoked. NEAR AI paused its cloud inference and promptly rotated credentials following the breach. The attack highlights the importance of scrutinizing the underlying infrastructure, as vulnerabilities in SGLang could potentially compromise the security of AI agents. It is crucial for developers to review and patch their systems, particularly focusing on the software running within the environment, as the isolation provided by such environments does not protect against the code within it.