Threat Actors Actively Targeting LLMs
Summary: GreyNoise's Ollama honeypot recorded 91,403 attack sessions between October 2025 and January 2026, split into two campaigns: an SSRF operation that spiked to 1,688 sessions in 48 hours over Christmas and showed a single JA4H signature in 99% of attacks across 62 IPs in 27 countries, and a separate probe starting December 28, 2025 in which two IPs generated 80,469 sessions against 73+ LLM model endpoints in eleven days. The second campaign tested OpenAI-compatible and Google Gemini API formats with deliberately innocuous queries, and its two IPs - 45.88.186.70 (AS210558, 1337 Services GmbH) with 49,955 sessions and 204.76.203.125 (AS51396, Pfcloud UG) with 30,514 - have histories of CVE exploitation including CVE-2025-55182 and CVE-2023-1389, with combined observations exceeding 4 million sensor hits.
What's compelling: Eighty thousand enumeration requests against exposed LLM endpoints represent investment, and GreyNoise assesses the operators are professional threat actors building target lists that feed a larger exploitation pipeline - meaning an exposed endpoint likely already sits on someone's map.
Read the original at GreyNoise Blog ↗
- Topics
- Vulnerabilities & Patching
- Mentions
- CVE-2023-1389CVE-2025-55182
- Related
- CVE-2025-55182: Meta React Server Components Remote Code Execution Vulnerability Known Exploited Vulnerabilities · Dec 05, 2025CVE-2023-1389: TP-Link Archer AX-21 Command Injection Vulnerability Known Exploited Vulnerabilities · May 01, 2023