Open-source AI “guardrails” stripped off as hackers eye exposed models, researchers say
Researchers identified over 175,000 internet-exposed open-source AI systems, with hundreds lacking safety guardrails and vulnerable to exploitation, according to SentinelOne and Censys. Nearly half of the hosts enabled tool-calling features, and about 30% were based in China, 20% in the U.S. Most ran versions of Meta’s Llama or Google’s Gemma models.