feat(airt): model-extraction & membership-inference attack generators#104
Merged
Conversation
Same missing-import bug as the ATLAS template (#100), in the separate `_build_agentic_imports`: generated agentic-attack scripts call get_generator (proxy-routing block) but never imported it, raising NameError at runtime. Add the import and a TestAgenticGeneration regression test (the agentic generator was previously untested).
…tors Adds generate_extraction_attack and generate_membership_attack to attack_runner. They build workflow scripts that construct a PredictionTargetSpec from a target's predict endpoint (or an environment challenge_url) and run the SDK's extraction (equation_solving/jacobian/copycat/knockoff) or membership (threshold/label_only) attacks under an Assessment, exporting airt spans to the platform. Verified: a generated knockoff-extraction workflow ran against a live fraud target (fidelity 0.93, 500 queries) and completed its assessment on the local platform.
Add generate_extraction_attack and generate_membership_attack @safe_tool wrappers so the AIRT agent (and TUI) can run the new traditional-ML privacy attacks, mirroring the existing generate_image_attack wrapper over the attack_runner generators.
Wire the boundary (numeric L2/Linf decision search) and text (token-flip) evasion attacks into the workflow runner and expose generate_evasion_attack as an agent tool. Bump capability to 1.9.0 and note traditional black-box ML coverage (extraction, membership inference, evasion) in the description.
Register the new named attacks in the generator dispatch maps: evasion (hopskipjump, simba, square, zoo, textfooler), extraction (activethief, distillation), and membership (lira, shadow_model, entropy, loss).
Update the attack_type parameter descriptions so the agent surfaces the new evasion (hopskipjump/simba/square/zoo/textfooler), extraction (activethief/ distillation), and membership (entropy/loss/shadow_model/lira) options.
rdheekonda
marked this pull request as ready for review
July 21, 2026 07:08
…mbership-attacks # Conflicts: # capabilities/ai-red-teaming/capability.yaml
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
Adds
generate_extraction_attackandgenerate_membership_attacktoattack_runner.py, so the AIRT capability can run the new traditional-ML privacy attacks (paired with SDK/platform support in dreadnode-tiger and target environments in security-tasks).PredictionTargetSpecfrom a target's predict endpoint (or an environmentchallenge_url) + optional API keyequation_solving/jacobian/copycat/knockoff) or membership (threshold/label_only) attack under anAssessment, exporting airt spansMETHODSdispatchValidation
A generated knockoff-extraction workflow ran against a live fraud target: SDK configured, pool fetched, attack ran (fidelity 0.93, 500 queries), Assessment registered + completed, spans exported to the local platform. Both generators produce self-compiling scripts.
Test plan
generate_extraction_attack+generate_membership_attackvia stdin dispatch