news

Prompt injection beat web agents in most attempts

Across 3,168 adversarial runs, no tested web agent stayed robust. Direct injection succeeded more than four times in five.

big-deal

What changed

A study called StakeBench, from teams at Nanyang Technological University, ST Engineering, IBM Research and the University of Illinois Urbana-Champaign, tested AI web agents across 3,168 adversarial runs and 264 benchmark cases in June 2026. Indirect prompt injection succeeded between 41.67 and 68.16 percent of the time. Direct injection exceeded 79 percent in every configuration tested.

Why you should care

If you are putting an agent in front of the open web on a client's behalf, these are the numbers you will be asked to answer for.

Read the source