PromptArmor finds open-weight models more prone to indirect prompt injection attacks
Overview
PromptArmor Threat Intelligence reports that open-weight models, including DeepSeek, Kimi, GLM and Qwen, failed 26.7% of 9,800 indirect prompt injection cyberattack trials, compared with 14.2% for proprietary models such as GPT, Claude and Gemini.
The firm says this is 88% higher susceptibility for the open-weight group.
The largest gaps appeared in two attack types: summarizing a URL that runs malware (49% for open-weight models versus 9% for proprietary models) and displaying a phishing login overlay (36% versus 5%).
Written by AI from the articles below · updated Oct 16, 4:25 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- PromptArmor Threat IntelligenceOfficialOpen-weight models show 88% higher susceptibility to indirect prompt injection attacks
AIPromptArmor Threat Intelligence reports that open-weight models, including DeepSeek, Kimi, GLM, and Qwen, were 88% more susceptible to real-world indirect prompt injection cyber attack chains than proprietary models across 9,800 trials. The open-weight models failed at 26.7% of attempts, compared with 14.2% for proprietary models such as GPT, Claude, and Gemini. Open-weight models showed the largest gaps on attacks that summarize a URL and run malware (49% vs. 9%) and that display a phishing login overlay (36% vs. 5%).
Heat trend
Not enough continuous observations to show a trend yet.