Repository navigation
feat(ai-red-teaming): expose 167 multimodal transforms to the TUI (v1.7.0) - #89
Merged
Merged
Conversation
…1.7.0)
Register the full SDK image/audio/video transform catalog in the multimodal
transform resolver so TUI/attack_runner attacks can apply any of them by name.
Grows the exposed media transforms from 26 to 131 (image 57, audio 52, video 22)
— the SDK-only transforms needing non-string args (interpolate_images,
overlay_image, frames_from_image_transform) are intentionally excluded.
- attack_runner.py: rebuild _MULTIMODAL_TRANSFORM_DEFS from per-modality name
lists mirroring dreadnode.transforms.{image,audio,video}
- tools/attacks.py: refresh the transforms param examples across modalities
- ai-red-teaming-agent.md: expand the technique->transform map (ImageNet-C,
SpecAugment, DolphinAttack, temporal video attacks)
- capability.yaml: minor version bump 1.6.10 -> 1.7.0; note 130+ multimodal
perturbations in the description
Expand the TUI-exposed multimodal transform registry from 131 to 166 (image 77, audio 61, video 28) to match the SDK round-2 additions (Albumentations/DSP gradient-free transforms). Refresh the description and agent counts to 160+.
rdheekonda
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Registers the full SDK image/audio/video transform catalog in the multimodal transform resolver so TUI /
attack_runnerattacks can apply any of them by name. Exposed media transforms grow from 26 → 131 (image 57, audio 52, video 22).This is the capability-side companion to dreadnode-tiger #1954, which adds the 130+ transform implementations to the SDK. Once both land and the capability publishes, TUI operators can request the new transforms directly (e.g. "apply
figstep_imageandfogto the image", "addultrasonic_shiftandfrequency_maskingto the audio").Changes
scripts/attack_runner.py— rebuild_MULTIMODAL_TRANSFORM_DEFSfrom per-modality name lists mirroringdreadnode.transforms.{image,audio,video}. SDK-only transforms needing non-string args (interpolate_images,overlay_image,frames_from_image_transform) are intentionally excluded.tools/attacks.py— refresh thetransformsparameter examples across all three modalities.agents/ai-red-teaming-agent.md— expand the attack technique → transform map with ImageNet-C corruptions, SpecAugment masking, DolphinAttack-style covert audio, and temporal/sparse-sampling video attacks; note 130+ multimodal transforms.capability.yaml— minor version bump 1.6.10 → 1.7.0; description notes 130+ image/audio/video perturbations spanning AugLy, ImageNet-C, SpecAugment, and published multimodal jailbreak families.Test plan
attack_runner.pyparses (AST);_MULTIMODAL_TRANSFORM_DEFSbuilds to 131 entries (image 57, audio 52, video 22)dreadnode.transforms.*)