36,769 Self‑Hosted AI Endpoints Exposed on the Internet, Only 2 % Protected by Authentication
What Happened — Researchers from Mysterium VPN scanned public‑internet indexes and identified 36,769 reachable AI endpoints (model servers, agent‑building platforms, vector stores). Only about 2 % returned an HTTP 401/403 response, meaning the overwhelming majority were accessible without any network‑layer authentication.
Why It Matters for Trust & Control Assurance
- Lack of authentication violates the core control objective of enforcing access controls for AI services; continuous discovery can surface this gap before it becomes a breach.
- Unauthenticated endpoints enable resource‑abuse and potential data leakage, eroding the evidentiary trail needed for audit readiness.
- Mapping these misconfigurations to a control‑assurance platform provides defensible evidence that satisfies multiple frameworks (e.g., NIST CSF 2.0).
Who Is Affected – Enterprises running self‑hosted large language models, AI platform vendors, cloud‑infrastructure teams, and any organization that deploys local model‑serving stacks.
Recommended Actions –
- Inventory every AI model‑serving, agent‑building, and vector‑store instance.
- Enforce network‑layer authentication (API keys, OAuth, mutual TLS) and restrict inbound traffic via firewalls or zero‑trust controls.
- Integrate automated endpoint discovery into your continuous control‑assurance program to generate audit‑ready evidence.
- Document configuration changes and access‑control settings for future assessments. Source: SecurityAffairs article
Technical Notes – The exposure stems from misconfiguration (absence of authentication) rather than a specific CVE. No vulnerability was actively exploited; researchers only queried a scanning index and verified HTTP response codes. Affected data could include proprietary model binaries, prompts, and any documents processed by the exposed services. Source: same as above