SGLang Flaw at NEAR AI Exposed 100 Accounts and 18 Agent Histories
Source
AI security news, week 46 of 2026
NEAR AI
Via NEAR AI
What happened
According to AI security news, week 46 of 2026, an attacker exploited a remote code execution flaw in SGLang. SGLang was the inference engine NEAR AI used to serve its models. The attacker obtained account details for approximately 100 users. The attacker also accessed chat histories, configuration settings, and logs belonging to 18 legacy agents. These were agents no longer actively in service, but their data had been retained on the platform. The incident exposes two distinct control failures: inadequate separation between the model-serving layer and stored customer data, and the continued retention of agent data beyond any operational need. No source confirmation was available that all exposed credentials or account details were live or exploitable at the time of access. The report does not specify whether the flaw affected only NEAR AI's deployment or other organizations running SGLang.
Why it matters
- ·Third-party inference engines such as SGLang sit between an organization's AI applications and its customer data. Most vendor risk programs do not assess them at the same depth as application software. A flaw at this layer can expose data across all services running on that infrastructure, making inference engine security a named gap in AI supply chain governance.
- ·The exposure of 18 legacy agents' histories shows that data minimization controls for AI agents are frequently absent. Retaining chat logs and configuration after agents are decommissioned creates residual liability with no operational benefit. Regulators applying frameworks such as the EU AI Act (Regulation (EU) 2024/1689) or the General Data Protection Regulation (GDPR) expect data to be deleted when no longer needed.
- ·This incident fits a pattern of AI platform breaches documented across the industry. Credential exposure and data leakage originate not in the AI model but in the surrounding serving infrastructure. Organizations that have inventoried their AI models but not the engines and middleware running them are carrying unassessed risk.
Stay ahead of stories like this
Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.
Recent issues
- The next quarter will test whether compliance programs can keep pace with AI systems that breach containment, reshape capital markets, and enable fraud at a scale no prior control framework anticipated.12 Nov
- State-linked attackers and criminal operators are actively weaponizing AI tools as entry points into enterprise systems, turning internal platforms and open-source frameworks into live attack vectors.5 Nov
Free every Thursday. Unsubscribe anytime.
Governance controls affected
What to do now
- ☐Ask your engineering or infrastructure team to list every inference engine or model-serving framework in use (such as SGLang, vLLM, or similar tools) and confirm that each is covered by your existing software vulnerability patch process.
- ☐Review data retention settings for all AI agents, including inactive or decommissioned ones, and confirm that chat histories, session logs, and configuration files are deleted on a defined schedule rather than held indefinitely.
- ☐Confirm that your model-serving infrastructure is isolated from broader customer account data stores, so that a flaw in the serving layer cannot reach user account details or agent histories.
- ☐Check whether your AI vendor contracts require the vendor to disclose vulnerabilities in underlying serving infrastructure, not only in the AI models themselves, and add that requirement to future procurement terms.
- ☐Verify that your AI incident response playbook covers breaches originating in third-party inference or serving components, not only in the AI model or application layer.
What to watch next
Organizations using managed AI platforms should monitor whether SGLang or other open-source inference engines release security patches. They should also watch whether any providers they rely on issue their own incident disclosures. The pattern of AI platform breaches, including nine AI platform breaches in October, suggests regulators will increasingly expect enterprises to assess the full AI serving stack, not only the model. Five Eyes Guidance on the Careful Adoption of Agentic AI Services already points toward supply chain controls for AI components. Enforcement attention to AI infrastructure security is growing alongside model-level scrutiny.