SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.06367 · 2026

WebSP-Eval: Evaluating Web Agents on Website Security and Privacy Tasks

Viswanathan Guruprasad, Asmit Ramesh, Basieem Nayak, Kassem Siddique, Fawaz

arXiv · PDF · Open in the Atlas

Code that ran

12 of 15 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, 1 violated it and 0 have no contract verdict; 10 failures of our instrument; 3 no recorded run), measured 2026-11-22; 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/2604.06367.json · badge: /api/ran/2604.06367.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
wi-pi/webspeval_code canonical (classifier label)
classifier basis: census
12 ran · 3 no recorded run
FunctionStatusWhere it lives
auto_eval_by_claude Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/judge/eval_binary_with_gt_text_claude.py
code served (permissive licence) · get_code("8eca7eb89de5f8fe")
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 tuple, 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.

import os
import json
import tempfile
import base64
from unittest.mock import Mock, MagicMock

# Create a temporary directory structure for testing
temp_dir = tempfile.mkdtemp()
process_dir = temp_dir

# Create interact_messages.json
interact_messages = [
    {
        "role": "system",
        "content": "You are a helpful assistant."
    },
    {
        "role": "user",
        "content": "Now given a task: Click the button labeled 'Submit' on the form. Please interact with the system to complete this task."
    },
    {
        "role": "assistant",
        "content": "I will help you click the Submit button."
    },
    {
        "role": "user",
        "content": "The action has been completed."
    }
]

with open(os.path.join(process_dir, 'interact_messages.json'), 'w') as f:
    json.dump(interact_messages, f)

# Create thoughts.json
thoughts_data = [
    {
        "iteration": 1,
        "thought": "I need to locate the Submit button on the form",
        "action": "Looking for button element with text 'Submit'"
    },
    {
        "iteration": 2,
        "thought": "Found the Submit button, now I'll click it",
        "action": "Clicking on the Submit button at coordinates (200, 150)"
    },
    {
        "iteration": 3,
        "thought": "Form submitted successfully",
        "action": "Waiting for confirmation"
    }
]

with open(os.path.join(process_dir, 'thoughts.json'), 'w') as f:
    json.dump(thoughts_data, f)

# Create dummy screenshot files
screenshot_data = b'\x89PNG\r\n\x1a\n\x00\x00\x00\rIHDR\x00\x00\x00\x01\x00\x00\x00\x01\x08\x06\x00\x00\x00\x1f\x15\xc4\x89\x00\x00\x00\nIDATx\x9cc\x00\x01\x00\x00\x05\x00\x01\r\n-\xb4\x00\x00\x00\x00IEND\xaeB`\x82'

for i in [1, 2, 3]:
    screenshot_path = os.path.join(process_dir, f'screenshot{i}.png')
    with open(screenshot_path, 'wb') as f:
        f.write(screenshot_data)

# Create mock Claude client
claude_client = Mock()
claude_client.messages = Mock()
claude_client.messages.create = MagicMock(return_value=Mock(
    content=[Mock(text="The agent successfully completed the task by clicking the Submit button. The evaluation shows that all steps were followed correctly.")]
))

# Define ground truth steps
ground_truth_steps = [
    "Navigate to the form page",
    "Locate the Submit button",
    "Click the Submit button",
    "Verify form submission confirmation"
]

ARGS = {
    'process_dir': process_dir,
    'claude_client': claude_client,
    'api_model': 'claude-3-5-sonnet-20241022',
    'system_prompt': 'You are an AI assistant evaluating the correctness of agent actions in completing web-based tasks.',
    'ground_truth_steps': ground_truth_steps,
    'has_timeout': False,
    'temperature': 1.0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple 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
check_for_api_errors Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/judge/eval_binary_with_gt_text_claude.py
code served (permissive licence) · get_code("2881b7afbbf74c18")
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 tuple, 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.

import os
import tempfile

# Create a temporary directory with an agent.log file
temp_dir = tempfile.mkdtemp()
log_file_path = os.path.join(temp_dir, 'agent.log')

# Write test content to the log file
with open(log_file_path, 'w') as f:
    f.write("Some agent execution log\n")
    f.write("Processing task...\n")
    f.write("Task execution time limit (180s) exceeded\n")
    f.write("API call error occurred\n")

ARGS = {
    'process_dir': temp_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple 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
encode_image_base64 Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/judge/eval_binary_with_gt_text_claude.py
code served (permissive licence) · get_code("66d017c4b55d2321")
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 base64
import tempfile
import os

# Create a temporary image file for testing
temp_dir = tempfile.mkdtemp()
image_path = os.path.join(temp_dir, "test_image.png")

# Create a minimal valid PNG file (1x1 pixel)
# This is the smallest valid PNG in bytes
png_bytes = bytes([
    0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A,  # PNG signature
    0x00, 0x00, 0x00, 0x0D, 0x49, 0x48, 0x44, 0x52,  # IHDR chunk
    0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,  # 1x1 dimensions
    0x08, 0x02, 0x00, 0x00, 0x00, 0x90, 0x77, 0x53,  # bit depth, color type
    0xDE, 0x00, 0x00, 0x00, 0x0C, 0x49, 0x44, 0x41,  # CRC, IDAT chunk
    0x54, 0x08, 0x99, 0x01, 0x01, 0x00, 0x00, 0xFE,  # compressed data
    0xFF, 0x00, 0x00, 0x00, 0x02, 0x00, 0x01, 0xE5,  # more data
    0x27, 0xDE, 0xFC, 0x00, 0x00, 0x00, 0x00, 0x49,  # CRC
    0x45, 0x4E, 0x44, 0xAE, 0x42, 0x60, 0x82          # IEND chunk
])

with open(image_path, 'wb') as f:
    f.write(png_bytes)

ARGS = {
    'image_path': image_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — 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
fetch_browser_info Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/utils_webarena.py
code served (permissive licence) · get_code("282e8a7fe50932e3")
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 dict. 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 typing import Any
from unittest.mock import Mock, MagicMock

# Create a mock browser object that implements the required methods
mock_browser = Mock()

# Mock the execute_cdp_cmd method to return a DOM snapshot structure
mock_browser.execute_cdp_cmd.return_value = {
    'documents': [
        {
            'layout': {
                'bounds': [
                    [10, 20, 1920, 1080],
                    [50, 60, 500, 300],
                    [100, 150, 800, 600]
                ]
            }
        }
    ]
}

# Mock the get_window_size method
mock_browser.get_window_size.return_value = {'width': 1920, 'height': 1080}

# Mock the execute_script method to return appropriate values
def execute_script_side_effect(script):
    if 'pageYOffset' in script:
        return 0.0
    elif 'pageXOffset' in script:
        return 0.0
    elif 'screen.width' in script:
        return 1920.0
    elif 'screen.height' in script:
        return 1080.0
    elif 'devicePixelRatio' in script:
        return 1.0
    return None

mock_browser.execute_script.side_effect = execute_script_side_effect

ARGS = {
    'browser': mock_browser
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
format_msg_claude Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/api_utils.py
code served (permissive licence) · get_code("a5524a41e66f9df4")
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 dict. 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 base64
from PIL import Image
import io

# Create a small base64-encoded PNG image for web_img_b64
img = Image.new('RGB', (10, 10), color='red')
img_bytes = io.BytesIO()
img.save(img_bytes, format='PNG')
img_bytes.seek(0)
web_img_b64 = base64.b64encode(img_bytes.read()).decode('utf-8')

# Define ARGS with all parameters
ARGS = {
    'it': 1,
    'init_msg': "Please help me navigate this webpage.",
    'pdf_obs': None,
    'warn_obs': "",
    'web_img_b64': web_img_b64,
    'web_text': "button: Click here\ninput: Search box",
    'tabs_info': {
        'tab_1': {
            'title': 'Google',
            'url': 'https://www.google.com',
            'is_current': True
        },
        'tab_2': {
            'title': 'GitHub',
            'url': 'https://www.github.com',
            'is_current': False
        }
    }
}
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 dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
format_msg_gpt5 Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/api_utils.py
code served (permissive licence) · get_code("76510fc830c652c8")
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 dict. 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 base64
import io
from PIL import Image

# Create a small base64 encoded image for web_img_b64
img = Image.new('RGB', (100, 100), color='red')
img_bytes = io.BytesIO()
img.save(img_bytes, format='PNG')
img_bytes.seek(0)
web_img_b64 = base64.b64encode(img_bytes.read()).decode('utf-8')

# Define arguments for format_msg_gpt5
ARGS = {
    'it': 1,
    'init_msg': 'You are a helpful assistant. Please help me interact with this webpage.',
    'pdf_obs': None,
    'warn_obs': '',
    'web_img_b64': web_img_b64,
    'web_text': '<button id="btn1">Click me</button>\n<input id="input1" type="text">Enter text here</input>\n<div id="div1">Some content</div>',
    'tabs_info': {
        'tab1': {
            'title': 'Google Search',
            'url': 'https://www.google.com',
            'is_current': True
        },
        'tab2': {
            'title': 'GitHub',
            'url': 'https://www.github.com',
            'is_current': False
        }
    }
}
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 dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
format_msg_text_only_gpt5 Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/api_utils.py
code served (permissive licence) · get_code("cf52e4eb5297bbbc")
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 dict. 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

# Setup for format_msg_text_only_gpt5

# Parameter: it (iteration number)
it = 1

# Parameter: init_msg (initial message string)
init_msg = "Please help me navigate this webpage."

# Parameter: pdf_obs (PDF observation string or empty)
pdf_obs = ""

# Parameter: warn_obs (warning observation string)
warn_obs = "Warning: Element not found"

# Parameter: ac_tree (accessibility tree as string)
ac_tree = """
RootWebArea "Example Page"
  navigation "Main Navigation"
    link "Home" <click>
    link "About" <click>
  main "Main Content"
    heading "Welcome"
    button "Submit" <click>
"""

# Parameter: tabs_info (dictionary of open tabs information)
tabs_info = {
    "tab_1": {
        "title": "Google Search",
        "url": "https://www.google.com",
        "is_current": True
    },
    "tab_2": {
        "title": "GitHub",
        "url": "https://www.github.com",
        "is_current": False
    }
}

ARGS = {
    "it": it,
    "init_msg": init_msg,
    "pdf_obs": pdf_obs,
    "warn_obs": warn_obs,
    "ac_tree": ac_tree,
    "tabs_info": tabs_info
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
get_bounding_client_rect Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/utils_webarena.py
code served (permissive licence) · get_code("040b32129bf67f60")
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 dict. 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 unittest.mock import Mock, MagicMock

# Create a mock browser object that implements the required methods
browser = Mock()

# Mock the execute_cdp_cmd method to return appropriate responses
def mock_execute_cdp_cmd(command, params):
    if command == 'DOM.resolveNode':
        return {
            'object': {
                'objectId': 'mock_object_id_12345'
            }
        }
    elif command == 'Runtime.callFunctionOn':
        return {
            'result': {
                'value': {
                    'x': 10.5,
                    'y': 20.3,
                    'width': 100.0,
                    'height': 50.0,
                    'top': 20.3,
                    'left': 10.5,
                    'bottom': 70.3,
                    'right': 110.5
                }
            }
        }
    return {}

browser.execute_cdp_cmd = mock_execute_cdp_cmd

backend_node_id = "42"

ARGS = {
    'browser': browser,
    'backend_node_id': backend_node_id
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
get_element_in_viewport_ratio Ran
ran, honoured the contract we drafted
wi-pi/webspeval_code/src/utils_webarena.py
code served (permissive licence) · get_code("0b27cebef7f8500d")
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

# Construct BrowserConfig TypedDict
config = {
    'win_top_bound': 0.0,
    'win_left_bound': 0.0,
    'win_width': 1920.0,
    'win_height': 1080.0,
    'win_right_bound': 1920.0,
    'win_lower_bound': 1080.0,
    'device_pixel_ratio': 1.0
}

# Define arguments for get_element_in_viewport_ratio
ARGS = {
    'elem_left_bound': 100.0,
    'elem_top_bound': 200.0,
    'width': 300.0,
    'height': 150.0,
    'config': config
}
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
get_tabs_info Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/agent/actions.py
code served (permissive licence) · get_code("602e1a57b115b9ef")
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 dict. 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 logging
from unittest.mock import Mock, MagicMock

# Create a mock driver/task object that simulates Selenium WebDriver behavior
driver_task = Mock()

# Set up window handles
handle1 = "CDwindow-1234567890"
handle2 = "CDwindow-0987654321"
handle3 = "CDwindow-1111111111"
window_handles = [handle1, handle2, handle3]

driver_task.window_handles = window_handles
driver_task.current_window_handle = handle2

# Create a side effect function for switch_to.window to track which window is active
current_window = [handle2]

def switch_window(handle):
    current_window[0] = handle

driver_task.switch_to.window = Mock(side_effect=switch_window)

# Mock the properties that change based on which window is active
def get_title():
    if current_window[0] == handle1:
        return "Google - Search"
    elif current_window[0] == handle2:
        return "GitHub"
    else:
        return "Stack Overflow"

def get_current_url():
    if current_window[0] == handle1:
        return "https://www.google.com/?q=python"
    elif current_window[0] == handle2:
        return "https://github.com/user/repo"
    else:
        return "https://stackoverflow.com/questions/tagged/python"

# Use PropertyMock to make title and current_url return different values
type(driver_task).title = property(lambda self: get_title())
type(driver_task).current_url = property(lambda self: get_current_url())

ARGS = {
    "driver_task": driver_task
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
load_json_file Ran
ran; our contract draft was wrong, not the code
wi-pi/webspeval_code/src/state_reset_check.py
code served (permissive licence) · get_code("53bdbb7cef5156de")
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 dict. 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 json
import tempfile
import os

# Create a temporary JSON file with valid content
temp_dir = tempfile.mkdtemp()
json_file_path = os.path.join(temp_dir, "test.json")

test_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "object"}
}

with open(json_file_path, 'w', encoding='utf-8') as f:
    json.dump(test_data, f)

ARGS = {
    "json_file": json_file_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
switch_to_latest_tab Ran
ran, violated the contract we drafted
wi-pi/webspeval_code/src/agent/actions.py
code served (permissive licence) · get_code("c0d2dc95bd1e83ea")
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 bool, 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 logging
from unittest.mock import Mock, MagicMock

# Create a mock driver_task object that has the necessary attributes and methods
driver_task = Mock()

# Set up window_handles to simulate browser tabs
# First call returns previous handles, second call returns current handles (with a new one added)
driver_task.window_handles = ["handle1", "handle2", "handle3"]

# Create a mock switch_to object with a window method
driver_task.switch_to = Mock()
driver_task.switch_to.window = Mock()

# Create previous_handles list (simulating tabs that existed before)
previous_handles = ["handle1", "handle2"]

ARGS = {
    "driver_task": driver_task,
    "previous_handles": previous_handles,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, 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
auto_eval_by_gpt5 Not yet run
unverified (harvested, no recorded run)
wi-pi/webspeval_code/judge/eval_binary_with_gt_text_gpt5.py
code served (permissive licence) · get_code("c8622f34c1208025")
How it ran

no invocation recorded.

create_fresh_profile Not yet run
unverified (harvested, no recorded run)
wi-pi/webspeval_code/src/agent/driver.py
code served (permissive licence) · get_code("5205d09c9c182848")
How it ran

no invocation recorded.

resize_image Not yet run
unverified (harvested, no recorded run)
wi-pi/webspeval_code/src/utils.py
code served (permissive licence) · get_code("e63c4f0834fe9c79")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 3 are also present in other papers whose code we hold (3 with at least one paper dated earlier, 0 the same month, 0 later) and 12 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
get_bounding_client_rect 1 other paper
1 dated earlier
2410.19609 (2024-10)
OpenWebVoyager: Building Multimodal Web Agents via Iterative Real-World Exploration, Feedback and Optimization
get_element_in_viewport_ratio 1 other paper
1 dated earlier
2410.19609 (2024-10)
OpenWebVoyager: Building Multimodal Web Agents via Iterative Real-World Exploration, Feedback and Optimization
resize_image 1 other paper
1 dated earlier
2410.19609 (2024-10)
OpenWebVoyager: Building Multimodal Web Agents via Iterative Real-World Exploration, Feedback and Optimization

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

Web agents automate browser tasks, ranging from simple form completion to complex workflows like ordering groceries. While current benchmarks evaluate general-purpose performance (e.g., We-bArena) or safety against malicious actions (e.g., SafeArena), no existing framework assesses an agent's ability to successfully execute user-facing website security and privacy tasks, such as managing cookie preferences, configuring privacy-sensitive account settings, or revoking inactive sessions. To address this gap, we introduce WebSP-Eval, an evaluation framework for measuring web agent performance on website security and privacy tasks. WebSP-Eval comprises 1) a manually crafted task dataset of 200 task instances across 28 websites; 2) a robust agentic system supporting account and initial state management across runs using a custom Google Chrome extension; and 3) an automated evaluator. We evaluate a total of 8 web agent instantiations using state-of-the-art multimodal large language models, conducting a fine-grained analysis across websites, task categories, and UI elements. Our evaluation reveals that current models suffer from limited autonomous exploration capabilities to reliably solve website security and privacy tasks, and struggle with specific task categories and websites. Crucially, we identify stateful UI elements are a primary reason for agent failure, with toggles causing more than 45% task failure across many models. Backbone LLM (VLM) Current Observation 𝒐 𝒕 A𝐜𝐭𝐢𝐨𝐧 𝒂 𝒕 Past Actions 𝒂 𝟏 , … , 𝒂 𝒕-𝟏 1. Go to "walmart.com" 2. Search for "3lb fuji apples" … "Add a 3lb bag of Fuji Apples to my Walmart Cart" User Instruction 𝐏 Context 𝒄 𝒕 Click "Add" button Past Observations 𝒐 𝟏 , … , 𝒐 𝒕-𝟏 1. Executes 𝒂 𝒕 in Environment E Added to context of next iteration 𝒄 𝒕+𝟏 Browser Automation Tool 2. Captures new observation 𝒐 𝒕+𝟏

For agents

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

get_harvested_code_for_paper("2604.06367")
get_code_for_paper("2604.06367")
have("2604.06367")

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

curl https://syntology.ai/api/ran/2604.06367.json

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

[![Syntology run record](https://syntology.ai/api/ran/2604.06367.svg)](https://syntology.ai/paper/2604.06367)

Connect an agent — have() is free.