A webpage can hide instructions for an artificial-intelligence search assistant. This indirect matters only if search puts the page in the assistant’s input; blocking suspicious pages can also discard useful information.
SearchInject scores pages for query relevance, signs of search manipulation, and instruction-like text, then holds back risky pages before the assistant sees them. Its cutoff is set using clean pages. Tests used separate queries and payloads, but attack types came from families seen during development.
Exposure to attack pages fell by 83.2–83.7% across tested search methods, with little change in clean rankings. Clean pages were held back more often than the calibration target allowed, and a separate generation test did not establish fewer successful attacks. The result supports exposure reduction for tested attack types, not general prevention.