Report: Internet-exposed Ollama AI servers widespread

Internet-exposed open-source Ollama servers used for downloading, executing, and managing large language models around the world totaled 175,000, nearly a third of which are in China, reports The Hacker News.

Tool calling functionality allowing code execution, API access, and real-time data retrieval was evident in 48% of the online Ollama hosts, highlighting the cybersecurity risk of such exposures, according to a joint analysis from SentinelOne SentinelLABS and Censys researchers.

"A text-generation endpoint can produce harmful content, but a tool-enabled endpoint can execute privileged operations. When combined with insufficient authentication and network exposure, this creates what we assess to be the highest-severity risk in the ecosystem," said researchers, who also suspected the systems' risk of being compromised in LLMjacking attacks.

Mounting edge deployment of LLMs was noted by researchers to warrant greater authentication, monitoring, and networking controls. Such findings come as exposed LLM endpoints were reported by Pillar Security to have been targeted by the new Operation Bizarre Bazaar LLMjacking campaign.