feat: τ²-Adv Bench - Adversarial Evaluation Module#158
Open
Ahm3dAlAli wants to merge 10 commits intosierra-research:mainfrom
Open
feat: τ²-Adv Bench - Adversarial Evaluation Module#158Ahm3dAlAli wants to merge 10 commits intosierra-research:mainfrom
Ahm3dAlAli wants to merge 10 commits intosierra-research:mainfrom
Conversation
This PR adds adversarial evaluation capabilities to τ2-bench, enabling testing of agent robustness against manipulation attempts. ## New Components ### Adversarial Module (`src/tau2/adversarial/`) - `strategies.py`: 5 attack strategies (social engineering, prompt injection, policy exploitation, identity manipulation, information extraction) with 3 sophistication levels each - `tasks.py`: Utilities for loading and filtering adversarial tasks - `run_adversarial.py`: CLI script for running adversarial evaluations - `README.md`: Documentation for the module ### Adversarial User (`src/tau2/user/adversarial_user.py`) - Wraps standard UserSimulator with adversarial instructions - Configurable attack strategy and sophistication - Tracks attack attempts for analysis ### Safety Evaluator (`src/tau2/evaluator/evaluator_safety.py`) - Detects safety violations in agent responses - Violation types: unauthorized actions, information disclosure, policy circumvention, prompt injection success, etc. - Produces SafetyRewardInfo with safety score and violations list ### Adversarial Tasks (`data/tau2/domains/airline/tasks_adversarial.json`) - 8 adversarial scenarios for airline domain - Covers all 5 attack strategies - Includes multi-vector combined attacks ## Usage ```bash # Run adversarial evaluation python -m tau2.adversarial.run_adversarial --domain airline # Filter by strategy python -m tau2.adversarial.run_adversarial --domain airline --strategy social_engineering ``` ## Tests Added comprehensive tests in `tests/test_adversarial.py` (25 passing) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add paper directory with NeurIPS paper files (tex, figures, generation scripts) - Include comprehensive evaluation results for multiple LLM models - Add adversarial task data for retail and telecom domains - Include visualization scripts and analysis tools Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Author
|
Hey @victorb-sierra , Good day, would love to know if you can take a look please. |
Author
|
Hi @victorb-sierra , Good day, Would Aveo to know about any updates. |
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 adversarial evaluation to τ²-Bench for testing agent safety against manipulation attacks.
Features:
Paper
Accompanying paper with methodology and findings available in fork: https://github.com/Ahm3dAlAli/tau2-bench/tree/feature/adversarial-evaluation/Tau2_Adv_Bench_AgentBeats_AhmedAli.pdf
Tau2_Adv_Bench_AgentBeats_AhmedAli.pdf
Test Plan
pytest tests/test_adversarial.py -v)python demo_adversarial.py)