Unpatched RCE in Open-Source AI Framework Leads to Data Breach Across Users and AI Agents
Written by AI from the linked story. Not reviewed by an editor. How this works
NEAR AI's inference server was compromised because it ran a vulnerable version of SGLang, an open-source AI inference engine with a known remote code execution flaw. The attacker leveraged this flaw to execute arbitrary code on the server, exposing personal data for roughly 100 users and sensitive logs—including API keys—from 18 AI agents. This incident illustrates the risk of deploying open-source dependencies without a disciplined process for tracking upstream vulnerability disclosures and applying patches promptly. The exposure of API keys in logs further compounded the breach, turning a single point of failure into a credential compromise event. Organizations building on open-source AI tooling must treat third-party inference engines with the same rigor as any other production dependency.
Tactical Insight
Immediate actions
- Identify and upgrade all instances of SGLang (and similar open-source inference engines) to the patched version immediately.
- Rotate all API keys and credentials that were accessible on affected servers, even if exposure is unconfirmed.
- Audit application and server logs to ensure secrets and credentials are never written in plaintext.
Long-term improvements
- Establish a formal software inventory (SBOM) for all open-source components used in AI/ML pipelines so vulnerabilities can be rapidly identified and scoped.
- Subscribe to upstream security advisories for every open-source dependency and define SLA-based patching windows for critical CVEs.
- Implement network segmentation to isolate inference servers from each other and from internal systems, limiting blast radius of any single server compromise.
Detection measures
- Deploy runtime monitoring and anomaly detection on inference servers to flag unexpected code execution or unusual process spawning.
- Implement secrets-scanning in CI/CD pipelines and log aggregation systems to detect and alert on credential exposure before an attacker finds them.
- Conduct regular vulnerability scans of internet-facing AI infrastructure, including third-party inference engines.