Hackers and other criminals can easily hijack computers running open-source large language models and use them for illicit activity, bypassing the safeguards built into major artificial intelligence platforms, researchers said on Thursday. The findings are based on a 293-day study conducted jointly by SentinelOne and Censys, and shared exclusively with Reuters. 

The research examined thousands of publicly accessible deployments of open-source LLMs and highlighted a broad range of potentially abusive use cases. According to the researchers, compromised systems could be directed to generate spam, phishing content, or disinformation while evading the security controls enforced by large AI providers. 

The deployments were also linked to activity involving hacking, hate speech, harassment, violent or graphic content, personal data theft, scams, fraud, and in some cases, child sexual abuse material. While thousands of open-source LLM variants are available, a significant share of internet-accessible deployments were based on Meta’s Llama models, Google DeepMind’s Gemma, and other widely used systems, the researchers said. 

They identified hundreds of instances in which safety guardrails had been deliberately removed. “AI industry conversations about security controls are ignoring this kind of surplus capacity that is clearly being utilized for all kinds of different stuff, some of it legitimate, some obviously criminal,” said Juan Andres Guerrero-Saade, executive director for intelligence and security research at SentinelOne. He compared the problem to an iceberg that remains largely unaccounted for across the industry and the open-source community. 

The study focused on models deployed using Ollama, a tool that allows users to run their own versions of large language models. Researchers were able to observe system prompts in about a quarter of the deployments analyzed and found that 7.5 percent of those prompts could potentially enable harmful behavior. 

Geographically, around 30 per cent of the observed hosts were located in China, with about 20 per cent based in the United States, the researchers said. Rachel Adams, chief executive of the Global Centre on AI Governance, said responsibility for downstream misuse becomes shared once open models are released.  “Labs are not responsible for every downstream misuse, but they retain an important duty of care to anticipate foreseeable harms, document risks, and provide mitigation tooling and guidance,” Adams said.  

A Meta spokesperson declined to comment on developer responsibility for downstream abuse but pointed to the company’s Llama Protection tools and Responsible Use Guide. Microsoft AI Red Team Lead Ram Shankar Siva Kumar said Microsoft believes open-source models play an important role but acknowledged the risks. 

“We are clear-eyed that open models, like all transformative technologies, can be misused by adversaries if released without appropriate safeguards,” he said. 

Microsoft conducts pre-release evaluations and monitors for emerging misuse patterns, Kumar added, noting that “responsible open innovation requires shared commitment across creators, deployers, researchers, and security teams.” 

Ollama, Google and Anthropic did not comment.