How reliable are prompt-injection detectors?
We tested seven detectors on 4,256 examples from five prompt-injection and jailbreak benchmarks. On one, the model with the highest F1 still flagged almost 59% of benign inputs.
Blog
Research and product notes from the team building Refractal.
All posts
-
How reliable are prompt-injection detectors?
We tested seven detectors on 4,256 examples from five prompt-injection and jailbreak benchmarks. On one, the model with the highest F1 still flagged almost 59% of benign inputs.
-
More coming soon
Research and product notes from the team building Refractal.
No posts on this topic yet.