SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.12814 · NeurIPS · 2024

Dissecting Adversarial Robustness of Multimodal LM Agents

Ruslan Salakhutdinov, Aditi Raghunathan, Daniel Fried, Chen Wu, Rishi Shah, Yu Koh

arXiv · PDF · Open in the Atlas

Code that ran

16 of 21 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 1 has no contract verdict; 14 failures of our instrument; 5 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2406.12814.json · badge: /api/ran/2406.12814.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
ChenWu98/agent-attack canonical (classifier label)
classifier basis: census
10 ran · 4 no recorded run
chenwu98/agent-attack canonical (classifier label)
classifier basis: classifier_2025
6 ran · 1 no recorded run
FunctionStatusWhere it lives
SpectrumSimulationAttack Ran
ran on a synthesized input
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("2eb813117afbc4ca")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, dtype family float; no output shape is stated. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import numpy as np

# Create a simple stand-in model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 8, kernel_size=3, padding=1)
        self.fc = nn.Linear(8 * 4 * 4, 10)
    
    def forward(self, x):
        x = self.conv1(x)
        x = torch.relu(x)
        x = torch.nn.functional.adaptive_avg_pool2d(x, (4, 4))
        x = x.view(x.size(0), -1)
        x = self.fc(x)
        return x

# Create two simple models for the list
model1 = SimpleModel()
model2 = SimpleModel()

INIT_ARGS = {
    'model': [model1, model2],
    'epsilon': 16 / 255,
    'norm': 'Linf'
}

# For the attack method, we need input images and labels
# The attack method signature is attack(self, *args, **kwargs)
# We'll create dummy inputs: batch_size=2, channels=3, height=4, width=4
ARGS = {
    'x': torch.randn(2, 3, 4, 4),
    'y': torch.tensor([0, 1])
}

METHOD = 'attack'
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
clamp Ran
ran; our contract draft was wrong, not the code
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("6007bd5812bfaa0f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a small tensor with values that need clamping
x = torch.tensor([[-0.5, 0.3, 1.2, 0.8],
                   [0.1, -0.2, 0.9, 1.5]], dtype=torch.float32)

ARGS = {
    "x": x,
    "min_value": 0,
    "max_value": 1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
create_prompt_with_image Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/models/claude.py
code served (permissive licence) · get_code("3da9b64f831dd75a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from PIL import Image
import io

# Create a small test image
test_image = Image.new('RGB', (10, 10), color='red')

# Define test prompt
test_prompt = "What is in this image?"

# Construct ARGS dict
ARGS = {
    'prompt': test_prompt,
    'image': test_image
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
create_prompt_with_image Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/models/gemini.py
code served (permissive licence) · get_code("2da643c5a93d2c26")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from PIL import Image
import io

# Create a simple test image
def create_test_image():
    # Create a small 10x10 RGB image
    img = Image.new('RGB', (10, 10), color='red')
    return img

# Set up arguments
test_image = create_test_image()
test_prompt = "This is a test prompt"

ARGS = {
    'prompt': test_prompt,
    'image': test_image
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
create_prompt_with_image Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/models/gpt4v.py
code served (permissive licence) · get_code("d4d849884c643a02")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from PIL import Image
import numpy as np

# Create a small test image
img_array = np.random.randint(0, 256, (32, 32, 3), dtype=np.uint8)
test_image = Image.fromarray(img_array, mode='RGB')

# Define arguments for create_prompt_with_image
ARGS = {
    'prompt': 'What is in this image?',
    'image': test_image,
    'detail': 'high'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
dct Ran
ran; our contract draft was wrong, not the code
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("4c3d07280a79397a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a small input tensor for the DCT function
# The function expects x to be a tensor with shape (..., N) where N is the last dimension
# Let's create a tensor with shape (2, 4) - batch size 2, signal length 4
x = torch.randn(2, 4)

ARGS = {
    'x': x,
    'norm': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
dct_2d Ran
ran; our contract draft was wrong, not the code
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("f9e84f203184602f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 3, dtype family float, shape (2, 4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a small 2D input tensor for dct_2d
# Shape: (batch=2, height=4, width=3)
x = torch.randn(2, 4, 3, dtype=torch.float32)

# Test with norm=None
ARGS = {
    'x': x,
    'norm': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 4, 3) — shape (2, 4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
edit_distance Ran
ran, honoured the contract we drafted
ChenWu98/agent-attack/agent_attack/attacks/utils.py
code served (permissive licence) · get_code("018efd26b7538a9c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# This function computes edit distance (Levenshtein distance) between two strings
# It takes two string parameters: s1 and s2
# The function compares characters and builds a DP table

s1 = "kitten"
s2 = "sitting"

ARGS = {
    "s1": s1,
    "s2": s2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
encode_image Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/models/claude.py
code served (permissive licence) · get_code("b594d1567e135be0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from PIL import Image
import io

# Create a small PIL Image object
img = Image.new('RGB', (4, 3), color='red')

ARGS = {
    'image': img
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
encode_image Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/models/gemini.py
code served (permissive licence) · get_code("72cb0591b722709a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bytes. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from io import BytesIO
from PIL import Image
import numpy as np

# Create a simple PIL Image object for testing
# The function calls image.save(buffered, format='PNG')
# So we need a PIL Image instance
image_array = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)
test_image = Image.fromarray(image_array, mode='RGB')

ARGS = {
    'image': test_image
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bytes — a bytes output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a bytes output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
evaluate_from_tensor Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/attacks/utils.py
code served (permissive licence) · get_code("a911ed75f6ec5cad")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockImageProcessor(nn.Module):
    def __init__(self):
        super().__init__()
    
    def image_processor_from_tensor(self, image):
        # Returns processed pixel values
        return torch.randn(2, 4)

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.image_processor = MockImageProcessor()
    
    def image_processor_from_tensor(self, image):
        # Process image and return pixel values
        return self.image_processor.image_processor_from_tensor(image)
    
    def generate_answer(self, pixel_values, inputs):
        # Generate text outputs based on pixel values and inputs
        # Return list of strings matching the batch size
        batch_size = pixel_values.shape[0]
        return ["hello world", "test output"]

model = MockModel()
image = torch.randn(2, 3, 224, 224)
inputs = ["question 1", "question 2"]
outputs = ["hello world", "test output"]

ARGS = {
    "model": model,
    "image": image,
    "inputs": inputs,
    "outputs": outputs
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
idct Ran
ran; our contract draft was wrong, not the code
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("d7527095d5399c56")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a small input tensor for the idct function
# The function expects X to be a tensor with shape (..., N) where N is the last dimension
# Let's create a tensor with shape (2, 4) - batch size 2, signal length 4
X = torch.randn(2, 4, dtype=torch.float32)

# The norm parameter can be None or 'ortho'
norm = None

ARGS = {
    'X': X,
    'norm': norm
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
idct_2d Ran
ran; our contract draft was wrong, not the code
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("7505f98253d5552a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 3, dtype family float, shape (2, 4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a small 2D input tensor for idct_2d
# Shape: (batch=2, height=4, width=3)
X = torch.randn(2, 4, 3, dtype=torch.float32)

ARGS = {
    "X": X,
    "norm": None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 4, 3) — shape (2, 4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
inplace_clamp Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("508ada73b3187beb")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a small tensor to clamp
x = torch.randn(2, 4)

ARGS = {
    "x": x,
    "min_value": 0.0,
    "max_value": 1.0,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_question_suffix Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/models/instructblip.py
code served (permissive licence) · get_code("2d7d16ad4a161f63")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

ARGS = {
    "question_suffix": "True or False question here? Yes or No answer expected."
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
scale_and_show_tensor Ran
ran; our contract draft was wrong, not the code
ChenWu98/agent-attack/agent_attack/util/image_handling.py
code served (permissive licence) · get_code("88fafc0c5b0a5f8d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Image. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from PIL import Image

torch.manual_seed(0)

# Create a small tensor with shape (1, 3, 4, 5) - batch=1, channels=3, height=4, width=5
# Values in range that will work with uint8 after scaling
x = torch.randn(1, 3, 4, 5) * 50 + 100

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
AdversarialInputAttacker Not yet run
unverified (harvested, no recorded run)
chenwu98/agent-attack/agent_attack/attacks/clip_attack.py
code served (permissive licence) · get_code("af1755102f0fb6e8")
How it ran

no invocation recorded.

dct1 Not yet run
unverified (harvested, no recorded run)
ChenWu98/agent-attack/agent_attack/attacks/clip_attack.py
pointer only (licence: MIT); the stored copy get_code serves, from gq-max/advdiffvlm, is pointer only (licence: NONE) · get_code("997a65deb3a03f4b")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from gq-max/advdiffvlm, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

evaluate_from_pil Not yet run
unverified (harvested, no recorded run)
ChenWu98/agent-attack/agent_attack/attacks/utils.py
code served (permissive licence) · get_code("fafceff1adc1b23c")
How it ran

no invocation recorded.

save_image Not yet run
unverified (harvested, no recorded run)
ChenWu98/agent-attack/agent_attack/util/image_handling.py
code served (permissive licence) · get_code("ddec91d5abd56bb7")
How it ran

no invocation recorded.

show_image Not yet run
unverified (harvested, no recorded run)
ChenWu98/agent-attack/agent_attack/util/image_handling.py
code served (permissive licence) · get_code("014944e169d043ef")
How it ran

no invocation recorded.

Also present in other papers' code

Of 21 samples lifted from this paper's linked repositories, 21 carry a function identity and 0 do not. Of the 21 distinct functions with an identity, 6 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 2 the same month, 2 later) and 15 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
dct1 4 other papers
3 dated earlier, 1 the same month
2306.12685 (2023-06)
Rethinking the Backward Propagation for Adversarial Transferability
2404.08255 (2024-04)
Practical Region-level Attack against Segment Anything Models
2404.10335 (2024-04)
Efficient Generation of Targeted and Transferable Adversarial Examples for Vision-Language Models Via Diffusion Models
and 1 more not named here
dct_2d
brick of this paper's assembly: data_transform
4 other papers
3 dated earlier, 1 later
2206.13397 (2022-06)
Generative Modelling With Inverse Heat Dissipation
2302.02693 (2023-02)
PatchDCT: Patch Refinement for High Quality Instance Segmentation
2306.00288 (2023-06)
Training-free Neural Architecture Search for RNNs and Transformers
and 1 more not named here
dct
brick of this paper's assembly: helper
3 other papers
2 dated earlier, 1 the same month
2306.12685 (2023-06)
Rethinking the Backward Propagation for Adversarial Transferability
2404.08255 (2024-04)
Practical Region-level Attack against Segment Anything Models
2406.05531 (2024-06)
Enhancing Adversarial Transferability via Information Bottleneck Constraints
idct_2d
brick of this paper's assembly: data_transform
2 other papers
1 dated earlier, 1 later
2302.02693 (2023-02)
PatchDCT: Patch Refinement for High Quality Instance Segmentation
2510.08273 (2025-10)
One Stone with Two Birds: A Null-Text-Null Frequency-Aware Diffusion Models for Text-Guided Image Inpainting
clamp
brick of this paper's assembly: helper
1 other paper
1 dated earlier
2402.02316 (2024-02)
Diffusion Models are Certifiably Robust Classifiers
inplace_clamp 1 other paper
1 dated earlier
2402.02316 (2024-02)
Diffusion Models are Certifiably Robust Classifiers

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

As language models (LMs) are used to build autonomous agents in real environments, ensuring their adversarial robustness becomes a critical challenge. Unlike chatbots, agents are compound systems with multiple components taking actions, which existing LMs safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation scripts in a realistic threat model on top of VisualWebArena, a real environment for web agents. To systematically examine the robustness of agents, we propose the Agent Robustness Evaluation (ARE) framework. ARE views the agent as a graph showing the flow of intermediate outputs between components and decomposes robustness as the flow of adversarial information on the graph. We find that we can successfully break latest agents that use black-box frontier LMs, including those that perform reflection and tree search. With imperceptible perturbations to a single image (less than 5% of total web page pixels), an attacker can hijack these agents to execute targeted adversarial goals with success rates up to 67%. We also use ARE to rigorously evaluate how the robustness changes as new components are added. We find that inference-time compute that typically improves benign performance can open up new vulnerabilities and harm robustness. An attacker can compromise the evaluator used by the reflexion agent and the value function of the tree search agent, which increases the attack success relatively by 15% and 20%. Our data and code for attacks, defenses, and evaluation are at github.com/ChenWu98/agent-attack.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2406.12814")
get_code_for_paper("2406.12814")
have("2406.12814")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2406.12814.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2406.12814.svg)](https://syntology.ai/paper/2406.12814)

Connect an agent — have() is free.