FEAT: Adversarial Models Blog Post - #2519
Conversation
Victor Valbuena (ValbuenaVC)
changed the title
| PyRIT helped us discover that we could improve the attack success rate of our adversarial chat target by more than 19 percentage points. | ||
|
|
||
| We use PyRIT during red teaming operations, where many automated techniques use an adversarial model for attack orchestration. We wanted better adversarial models for PyRIT, so we built a tool to automate that comparison process. This became PyRIT's `AdversarialBenchmark` scenario and turned model selection from an intuition-driven choice into a repeatable evaluation. We are scaling this evaluation through CI/CD to continuously discover effective models for automating red teaming. When we ran it against in-house models, we discovered surprising patterns in attack performance across techniques, which this post explores. | ||
|
|
There was a problem hiding this comment.
Another thing I might add to the first paragraph is a recommendation; "Grok is a decent adversarial model to start with".
One of the biggest hurdles to getting started is selecting an adversarial model, so writing this down I think can be helpful
There was a problem hiding this comment.
I may disagree on making a recommendation to use a specific model. The blog post ends with noting that the "best" model depends on technique, objectives, scoring, etc. So, I feel like pointing to a provider even as a good place to start with is misleading. Thoughts?
There was a problem hiding this comment.
I still think it's nice to say it's decent, even if we caveat it. I know it will change, but I want to make things as easy as we can for folks to get started is a good thing
Richard Lundeen (rlundeen2)
left a comment
Richard Lundeen (rlundeen2)
left a comment
There was a problem hiding this comment.
Looks good! Approved with smallish things
Victor Valbuena (ValbuenaVC)
changed the title
Co-authored-by: hannahwestra25 <hannahwestra@microsoft.com>
Co-authored-by: hannahwestra25 <hannahwestra@microsoft.com>
Victor Valbuena (ValbuenaVC)
changed the title
3 participants
Description
Blog post for the adversarial model benchmark (
AdversarialBenchmark) in PyRIT. The blog post explains what adversarial models are, why they matter in PyRIT, and what happened when we created and ran a benchmarking scenario on them. Small datasets and graphs are attached in this PR.Tests and Documentation
N/A. All changes are under
/doc/blog.