LoopLeap Lab

NeurIPS 2026 · Evaluations & Datasets Track

RedPhys: An Extensible Red-Teaming Platform for Physical-World Models and Agents

Configure physical-world tasks, test methods, defenses, and victim models or agents. Compare functional robustness and collect reusable test datasets for future alignment.

Yifei Zhao · Qian Lou · Mengxin Zheng
LoopLeap Lab · University of Central Florida (UCF)

12attacks
3defenses
13PhyModels
  1. watching the victim…
An extensible platform

Configure the test. Reuse the evidence.

Adversarial stress testing is the main focus. The same interfaces also support non-adversarial input perturbations and obfuscations without an attack objective.

RedPhys connects physical tasks, test methods, protection, and PhyModels or PhyAgents to shared evaluation, robustness reports, and datasets for future alignment.
The paper's overview. Dashed “Add” slots indicate extensible components; outputs support comparison, reproducibility, and future alignment.

PhyModels

Physical-world models perform grounded perception, reasoning, prediction, or decision-making about physical scenes.

PhyAgents

Physical-world agents use these models with planning or tool use to carry out grounded tasks.

Scope of the current paper. The study focuses on vision-language PhyModels and tool-using PhyAgents built on vision-language backbones. Evaluation uses scene images, questions, and task-level responses. These terms describe physical-world task roles across model architectures.

Why RedPhys

Harmful text vs. wrong actions

Chatbot red-teaming benchmarksHarmBench · AgentHarm · AdversarialLLM · OpenRTevaluates harmful content generation
Fixed harmful behavior
cybercrime
drawn from a fixed list
Jailbreak attack

adversarial text added to the prompt
Chatbot reply
I can’t help with that
refused
RedPhysPhyModels and PhyAgentsevaluates physically grounded functional failures
Real scene + questionThe real scene
Where is the pedestrian with the cart?
Scene-specific targetThe same scene edited to show the target
target answer: clear of the path
AttackThe original scene perturbed toward the target
scene perturbed toward the target
PhyModel answer
crossing the car's path ✓clear of the pathpedestrian collision risk

Functional safety

Failures are wrong actions and judgments in grounded tasks, not harmful text.

Scene-specific targets

A scenario agent builds each target from its scene, the same for every method.

One protocol

One dataset, two metrics and a black-box threat model for attacks and defenses.

Real scenes

Driving, manipulation, grasping, physics, egocentric, diagnostics and scene QA.

How it works

From physical-world scenes to robustness evidence

Physical domains
    Task familiesperceptionpredictionreasoningevaluationplanning
    Failure modesaction predictiontask successspatial reasoningphysical plausibilityembodied planning
    TargetsA scenario agent reads each scene and derives a target answer and a target image.
    Attack12 methods build each case once, offline, on surrogate models or generators.
    Defense3 defenses, one preprocessing step before the victim, no retraining.
    Victim13 PhyModels answer. We report ASR and clean accuracy.
    Library

    Attacks and defenses

    Results

    12 attacks against 13 PhyModels

    832 physical-world cases, one-shot black-box transfer. ASR is the attack success rate.

    Average ASR per attack% over 13 PhyModels
    Per modelclick a column to sort
    Clean accuracy vs. average ASR13 PhyModels

    Models that answer clean images less accurately tend to be easier to attack, though not deterministically.

    DefensesASR %, avg of 9 PhyModels

    Lower is better. BlueSuffix cuts PromptInject and FreqPure cuts FOA. PAD changes almost nothing, and no defense lowers FigStep.

    Quick start

    From clone to report

    
        
    Roadmap

    What is delivered, and what comes next

    The toolkit and the benchmark are out. An automatic platform is in progress, then agents that research new attacks and defenses.

    delivered

    Attacks and defenses

    One library to run, defend against and score red-teaming methods.

    • 12 attacks and 3 defenses, one API
    • Paper configs, install script and CLI
    delivered

    Agentic pipeline and benchmark

    Agents build a target for each scene; the benchmark scores 13 PhyModels.

    • Scenario agent: a target per scene
    • Hazard, physics, medical and NIPS 2017 sets
    in progress

    Automatic red‑teaming platform

    Fully automatic. The user picks a model or a use case; agents do the rest.

    • Input: a model or a use case
    • Output: a weakness report
    See the plan →
    future

    Agentic evolve

    Agents design new attacks and defenses and test them.

    • Attack and defense adapt round by round
    • Tested on held-out victims
    See it run →
    Automatic red‑teaming platform
    Agentic evolve
    Co-evolution loop
    Attack
    Defense
    round01of 08
      
    Across rounds attack success defense strength illustrative
    Attack agentthinking
    
            
    Open leaderboardillustrative
    #victimrobustness under evolving attacks
    Cite

    Reference RedPhys

    Yifei Zhao, Qian Lou, and Mengxin Zheng · LoopLeap Lab

    RedPhys is accepted at NeurIPS 2026 (Evaluations & Datasets Track). The citation below reflects the accepted conference paper; the arXiv version remains available. Download BibTeX.

    BibTeX
    @inproceedings{zhao2026redphys,
      title={{RedPhys}: An Extensible Red-Teaming Platform for Physical-World Models and Agents},
      author={Zhao, Yifei and Lou, Qian and Zheng, Mengxin},
      booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
      year={2026},
      note={Accepted, Evaluations and Datasets Track},
      doi={10.48550/arXiv.2606.23892},
      url={https://arxiv.org/abs/2606.23892}
    }