Threat actors probe LLM endpoints and misconfigured proxies

What happened

Security researchers and threat-intel teams observed a widespread reconnaissance campaign probing LLM endpoints and misconfigured proxies from Jan 8–12, 2026. GreyNoise flagged active targeting on Jan 8. SC Media reported honeypots detecting mass scans on Jan 9. BleepingComputer detailed attackers abusing misconfigured proxies to reach paid LLM services on Jan 9. DarkReading described two separate campaigns on Jan 12. All reports describe automated, high-volume probes rather than single-target intrusions.

Technical context

Attackers scanned broadly for exposed model endpoints and improperly secured reverse proxies. They attempted to use misconfigured proxies to bypass authentication, elevate access, or harvest API keys and session tokens. BleepingComputer documents direct attempts to route requests through misconfigured proxies to reach paid model instances. DarkReading distinguishes discovery and scan activity from follow-on exploitation. GreyNoise telemetry and SC Media honeypots show consistent scanning signatures and rapid probe cycles, suggesting tool reuse and commoditization of LLM-targeting scanners.

Implications and next steps

Reconnaissance often precedes credential theft, model abuse (for example, prompt-injection or paid-service fraud), or supply-chain probing. Defenders should inventory proxy and gateway configurations, enforce strong authentication and per-key rate limits on LLM endpoints, rotate any exposed API keys, and deploy honeypots and traffic baselining to detect anomalous probe patterns.

Why It Matters

• Inventory and lock down reverse proxies and API gateways now — misconfigurations are being actively scanned and abused.
• Enforce strong authentication, per-key rate limits, and comprehensive logging on LLM endpoints to reduce account and service abuse and speed incident response.
• Deploy honeypots and baseline network/model traffic to detect automated scanning signatures and tooling reuse before exploitation.
• Rotate exposed API keys immediately and monitor for anomalous model usage (unusual volume, geographic regions, or prompt patterns).