Learn about this first
PromptArmor released an evaluation reporting that, across 9800 trials, open-weight models (DeepSeek, Kimi, GLM, Qwen) had a 26.7% susceptibility rate to real-world indirect prompt injection attack chains, compared with 14.2% for closed-source models (GPT, Claude, Gemini)—an 88% higher susceptibility rate for the former.
Restating the findings on social media, PromptArmor said it tested the “most widespread attacks” and claimed that DeepSeek, Kimi, GLM, and Qwen were 88% more susceptible to cyberattacks than GPT, Claude, and Gemini. This wording frames the findings as suggesting that Chinese models are more likely to leak data.
These results come from PromptArmor’s own evaluation and reflect the relative performance of the two groups of models under its test conditions. They do not constitute independent verification of the security of specific models.
AI-generated from reports · updated 19 days agoLatest developments
Are Chinese Models More Likely to Leak Data?
Event progress1
PromptArmor Releases an Evaluation of Open-Source Models’ Susceptibility to Attacks
PromptArmor: Threat Intelligence: PromptArmor Study Finds Open-Weight Models Have an 88% Higher Susceptibility Rate to Indirect Prompt Injection Attacks