AI Agents Vulnerable to Prompt Injection Attacks, Researchers Find 79% Success Rate

According to researchers from Nanyang Technological University, ST Engineering, IBM Research, and the University of Illinois Urbana-Champaign, AI agents powered by GPT-5 and Gemini cannot consistently resist prompt injection attacks, a study published on Thursday found.

In 3,168 attack simulations, direct prompt injection attacks succeeded more than 79% of the time, while indirect attacks embedded in web content achieved success rates between 41.67% and 68.16%. The researchers developed StakeBench, a benchmark to test AI agent responses to such attacks in realistic online environments, and noted that prompt injection remains a critical vulnerability as AI agents become mainstream for web browsing, research, shopping, and cryptocurrency trading.