SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2409.11295 · ICLR · 2025

EIA: Environmental Injection Attack on Generalist Web Agents for Privacy Leakage

Jiawei Zhang, Bo Li, Chejian Xu, Mintong Kang, Lingbo Mo, Zeyi Liao, ♢ Chaowei, Xiao Yuan, ♢ Huan

arXiv · PDF · Open in the Atlas

Code that ran

13 of 14 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; 11 failures of our instrument; 1 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/2409.11295.json · badge: /api/ran/2409.11295.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
osu-nlp-group/eia_against_webagent canonical (classifier label)
classifier basis: census
13 ran · 1 no recorded run
FunctionStatusWhere it lives
clean_text Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/dom_utils.py
code served (permissive licence) · get_code("0e07e2733eab1746")
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 re

def clean_text(text):
    if text is None:
        return ''
    text = re.sub('\\s+', ' ', text)
    return text.strip()

ARGS = {
    'text': '  hello   world  \n  foo  bar  '
}
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
extract_elements_by_ids Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/image_utils.py
code served (permissive licence) · get_code("b64eb9b4e7d05784")
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.

import torch
import numpy as np

# Create sample elements with 'backend_node_id' keys
all_elements = [
    {'backend_node_id': 1, 'name': 'element_1', 'value': 10},
    {'backend_node_id': 2, 'name': 'element_2', 'value': 20},
    {'backend_node_id': 3, 'name': 'element_3', 'value': 30},
    {'backend_node_id': 4, 'name': 'element_4', 'value': 40},
    {'backend_node_id': 5, 'name': 'element_5', 'value': 50},
]

# Create a list of ids to extract (order matters - should be preserved in output)
ids = [3, 1, 5]

ARGS = {
    'all_elements': all_elements,
    'ids': ids,
}
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 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
extract_topk_elements Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/image_utils.py
code served (permissive licence) · get_code("9fb9dab3374bb263")
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.

import copy

ARGS = {
    'all_elements': [
        {'rank': 0, 'score': 0.9, 'id': 1},
        {'rank': 1, 'score': 0.8, 'id': 2},
        {'rank': 2, 'score': 0.7, 'id': 3},
        {'rank': 5, 'score': 0.5, 'id': 4},
        {'rank': 10, 'score': 0.3, 'id': 5},
    ],
    'k': 3,
}
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 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
generate_new_query_prompt Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/format_prompt_utils.py
code served (permissive licence) · get_code("3db599f0f49fee4f")
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.

import torch
import numpy as np

torch.manual_seed(0)

ARGS = {
    'system_prompt': 'You are a helpful assistant.',
    'task': 'Navigate to the target location and complete the mission.',
    'previous_actions': [
        'Moved forward 5 steps',
        'Turned left 90 degrees',
        'Picked up item from ground'
    ],
    'question_description': 'What is your current understanding of the environment and what is your next action?'
}
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 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
generate_query_prompt Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/dy_human_dynamic.py
code served (permissive licence) · get_code("dcac82655df6139c")
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
import numpy as np

sys_prompt = 'Imagine that you are imitating human doing web navigation for a task step by step. At each stage, you can see the webpage like human by a given screenshot, and know the previous action history indicating what have done before the current stage. You can click an element with the mouse, select an option, or type some text with the keyboard.\n\nYou are asked to complete the following task: '

question_description_new_human_eval = "Below the screenshot is the webpage you see. Follow the following guidance to think step by step before clearly outlining the next action step at current stage that user would like to take(Not all the subsequent actions):\n\n(Current Webpage Identification)Firstly, think about what the current webpage is.\n\n(Previous Action Analysis)Secondly, combining with the screenshot, analyze each step of the previous action history and the intention of each of them one by one. Particularly, pay more attention to the last step, which may be more related to what you should do now as the next step.\n\n(Screenshot Details Analysis) Closely examine the screenshot to check the status of every part of the webpage, to understand what you can operate with, and what have been set or completed. You should closely examine the details of the screenshot to see what steps has been completed by previous actions even though you are given the textual previous actions. Because some effect of previous actions may not be clearly and sufficiently recorded by the textual history, you should really closely evaluate the status of every part of the webpage to understand what have done.\n\n(Intended Action Based on Webpage and Analysis)Then, based on your analysis, in conjunction with human's web browsing habits and the logic of web design, conclude which element in the webpage will users operate with as the first next target element, where the element is located, and what the corresponding operation is. Please clearly specify which element to operate with. Closely examine the screenshot to provide its detailed location, description, textual content(if it has) and so on to help identify your answer. If there are multiple elements that are similar to your target element, use more precise description to make sure people can distinguish your target element from them through your answer."

ARGS = {
    'system_prompt': sys_prompt,
    'task': 'Find and book a hotel in Paris for 2 nights',
    'previous_actions': [
        'Action 1: Clicked on the search bar at the top of the page',
        'Action 2: Typed "Paris hotels" in the search box',
        'Action 3: Pressed Enter to search'
    ],
    'question_description': question_description_new_human_eval
}
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 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
generate_query_prompt Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/format_prompt_utils.py
code served (permissive licence) · get_code("061acc9fbe79bb1c")
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

# Define ARGS dict with all parameters for generate_query_prompt
ARGS = {
    'system_prompt': 'You are a helpful AI assistant.',
    'task': 'Complete the following task: navigate to the red object and pick it up.',
    'previous_actions': [
        'Action 1: Moved forward 2 steps',
        'Action 2: Turned left 90 degrees',
        'Action 3: Observed red object at distance 5 meters'
    ],
    'question_description': 'What should be the next action to take?'
}
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
get_descendants Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/dom_utils.py
code served (permissive licence) · get_code("56fc0416b2cf42ae")
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.

class Node:
    def __init__(self, children=None):
        self.children = children if children is not None else []
    
    def __iter__(self):
        return iter(self.children)


# Build a small tree structure for testing
leaf1 = Node([])
leaf2 = Node([])
leaf3 = Node([])
middle1 = Node([leaf1, leaf2])
middle2 = Node([leaf3])
root = Node([middle1, middle2])

ARGS = {
    "node": root,
    "max_depth": 2,
    "current_depth": 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 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
init_cdp_session Ran
ran on a synthesized input
osu-nlp-group/eia_against_webagent/SeeAct/src/dy_human_dynamic.py
code served (permissive licence) · get_code("b6a99db34d8f1684")
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 coroutine. 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 asyncio
from unittest.mock import AsyncMock, MagicMock


# Create mock objects that simulate the Playwright page and CDP session
mock_cdp_session = AsyncMock()
mock_cdp_session.send = AsyncMock(return_value=None)

mock_context = AsyncMock()
mock_context.new_cdp_session = AsyncMock(return_value=mock_cdp_session)

mock_page = AsyncMock()
mock_page.context = mock_context

ARGS = {
    'page': mock_page
}
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 coroutine — a coroutine output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a coroutine 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
normal_new_context_async Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/demo_utils/browser_helper.py
code served (permissive licence) · get_code("f9d1a77f72c2e2f9")
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 coroutine. 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 random
import asyncio
from unittest.mock import AsyncMock, MagicMock

# Create a mock browser object that supports the async context manager protocol
mock_browser = AsyncMock()

# Create a mock context that will be returned by browser.new_context
mock_context = AsyncMock()
mock_browser.new_context = AsyncMock(return_value=mock_context)

# Define ARGS with all parameters for normal_new_context_async
ARGS = {
    'browser': mock_browser,
    'storage_state': None,
    'har_path': None,
    'video_path': None,
    'tracing': False,
    'trace_screenshots': False,
    'trace_snapshots': False,
    'trace_sources': False,
    'locale': None,
    'geolocation': None,
    'user_agent': 'Mozilla/5.0 (Windows NT 10.0; WOW64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/109.0.0.0 Safari/537.36',
    'viewport': {'width': 1280, 'height': 720}
}
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 coroutine — a coroutine output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a coroutine 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
process_attack_type_subtypes Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/injection/inject_malicious_form.py
code served (permissive licence) · get_code("c0e8041ef4fac1b2")
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

torch.manual_seed(0)

# The function process_attack_type_subtypes takes two string parameters
# and returns a concatenated string based on the attack_type value.
# No complex objects are needed.

ARGS = {
    'attack_type': 'dos',
    'attack_subtype': 'syn_flood'
}
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
remove_extra_eol Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/dom_utils.py
code served (permissive licence) · get_code("d6c9b5270beb73e5")
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 re

def remove_extra_eol(text):
    text = text.replace('\n', ' ')
    return re.sub('\\s{2,}', ' ', text)

ARGS = {
    'text': 'Hello\nworld  \n  this is\n\n  a test'
}
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
similar Ran
ran, honoured the contract we drafted
osu-nlp-group/eia_against_webagent/SeeAct/src/dy_human_dynamic.py
code served (permissive licence) · get_code("5e4b02e5bc753357")
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 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.

from difflib import SequenceMatcher

# The similar function takes two arguments (a, b) and compares them using SequenceMatcher
# SequenceMatcher works with any sequence type (strings, lists, etc.)
# We'll use strings as they're the most common use case

a = "hello world"
b = "hello there"

ARGS = {
    "a": a,
    "b": b,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 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
write_html_file Ran
ran; our contract draft was wrong, not the code
osu-nlp-group/eia_against_webagent/injection/inject_malicious_form.py
code served (permissive licence) · get_code("ce51ce402f71333e")
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 tempfile
import os

# Create a temporary HTML file for testing
temp_dir = tempfile.mkdtemp()
html_file_path = os.path.join(temp_dir, "test.html")

# Write a sample HTML file with nested divs and outer_html marker
html_content = """<!DOCTYPE html>
<html>
<head><title>Test</title></head>
<body>
<div id="outer">
    <div id="inner1">
        <p>Content 1</p>
    </div>
    <div id="inner2">
        <p>Content 2</p>
    </div>
</div>
</body>
</html>"""

with open(html_file_path, 'w', encoding='utf-8') as f:
    f.write(html_content)

# Define arguments for the function
ARGS = {
    'html_file_path': html_file_path,
    'outer_html': '<div id="outer">',
    'x': 0,
    'direction': 'top',
    'form_html': '<form><input type="text"></form>'
}
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
generate_prompt Not yet run
unverified (harvested, no recorded run)
osu-nlp-group/eia_against_webagent/SeeAct/src/data_utils/prompts.py
code served (permissive licence) · get_code("431150ea1b184bf1")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 6 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 0 the same month, 1 later) and 8 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
similar 6 other papers
5 dated earlier, 1 later
2301.11305 (2023-01)
DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature
2305.00909 (2023-05)
Outline, Then Details: Syntactically Guided Coarse-To-Fine Code Generation
2310.05914 (2023-10)
NEFTune: Noisy Embeddings Improve Instruction Finetuning
and 3 more not named here
get_descendants 3 other papers
3 dated earlier
2401.01614 (2024-01)
GPT-4V(ision) is a Generalist Web Agent, if Grounded
2402.05930 (2024-02)
WebLINX: Real-World Website Navigation with Multi-Turn Dialogue
2402.15057 (2024-02)
On the Multi-turn Instruction Following for Conversational Web Agents
clean_text 2 other papers
2 dated earlier
2401.01614 (2024-01)
GPT-4V(ision) is a Generalist Web Agent, if Grounded
2402.15057 (2024-02)
On the Multi-turn Instruction Following for Conversational Web Agents
generate_new_query_prompt 1 other paper
1 dated earlier
2401.01614 (2024-01)
GPT-4V(ision) is a Generalist Web Agent, if Grounded
generate_query_prompt 1 other paper
1 dated earlier
2401.01614 (2024-01)
GPT-4V(ision) is a Generalist Web Agent, if Grounded
remove_extra_eol 1 other paper
1 dated earlier
2401.01614 (2024-01)
GPT-4V(ision) is a Generalist Web Agent, if Grounded

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

Recently, generalist web agents have demonstrated remarkable potential in autonomously completing a wide range of tasks on real websites, significantly boosting human productivity. However, web tasks, such as booking flights, usually involve users' personally identifiable information (PII), which may be exposed to potential privacy risks if web agents accidentally interact with compromised websites-a scenario that remains largely unexplored in the literature. In this work, we narrow this gap by conducting the first study on the privacy risks of generalist web agents in adversarial environments. First, we present a realistic threat model for attacks on the website, where we consider two adversarial targets: stealing users' specific PII or the entire user request. Then, we propose a novel attack method, termed Environmental Injection Attack (EIA). EIA injects malicious content designed to adapt well to environments where the agents operate and our work instantiates EIA specifically for privacy scenarios in web environments. We collect 177 action steps that involve diverse PII categories on realistic websites from the Mind2Web dataset, and conduct experiments using one of the most capable generalist web agent frameworks to date. The results demonstrate that EIA achieves up to 70% attack success rate (ASR) in stealing users' specific PII and 16% ASR in stealing a full user request at an action step. Additionally, by evaluating the detectability and testing defensive system prompts, we indicate that EIA is challenging to detect and mitigate. Notably, attacks that are not well adapted for a webpage can be detected through careful human inspection, leading to our discussion about the trade-off between security and autonomy. However, extra attackers' efforts can make EIA seamlessly adapted, rendering such human supervision ineffective. Thus, we further discuss the implications on defenses at the pre-and post-deployment stages of the websites without relying on human supervision and call for more advanced defense strategies.

For agents

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

get_harvested_code_for_paper("2409.11295")
get_code_for_paper("2409.11295")
have("2409.11295")

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

curl https://syntology.ai/api/ran/2409.11295.json

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

[![Syntology run record](https://syntology.ai/api/ran/2409.11295.svg)](https://syntology.ai/paper/2409.11295)

Connect an agent — have() is free.