OpenRouter tests prompt injection defenses across 25 attacks and four models
Overview
OpenRouter tested 25 prompt injection attacks, 13 written by its team and 12 by other models, against four current models with each defense switched on and off.
In its blog post, the company reports that its Armor tool-call checker cut successful attack emails from 73 of 1,440 attempts to 0, while nearly all legitimate emails still went through.
Labeling untrusted documents as data reduced attack emails from 60 to 13 of 720 attempts. OpenRouter also found that one model sent 58 of 60 attack emails when no defenses were enabled.
Written by AI from the articles below · updated Oct 28, 9:07 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- OpenRouter BlogOfficialHow to prevent prompt injection, tested across 13 attacks and four models
AIOpenRouter tested 25 prompt injection attacks, 13 written by its team and 12 by other models, against four current models with each defense switched on and off. With its Armor tool-call checker enabled, attack emails dropped from 73 of 1,440 attempts to 0, while nearly all legitimate emails still went through. Labeling untrusted documents as data cut attack emails from 60 to 13 of 720 attempts, and the team found one model sent 58 of the 60 attack emails with no defenses.
Heat Trend
Not enough continuous observations to show a trend yet.