SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2407.20224 · 2024

Can Editing LLMs Inject Harm?

arXiv · PDF · Open in the Atlas

Code that ran

10 of 13 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; 0 honoured the contract we drafted, 0 violated it and 5 have no contract verdict; 5 failures of our instrument; 3 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/2407.20224.json · badge: /api/ran/2407.20224.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
llm-editing/editing-attack community (classifier label)
classifier basis not recorded on the link
10 ran · 3 no recorded run
FunctionStatusWhere it lives
binary_log_probs Ran
ran; our contract draft was wrong, not the code
llm-editing/editing-attack/code/easyeditor/trainer/losses.py
pointer only (licence: MIT); the stored copy get_code serves, from MemEIC/MemEIC, is pointer only (licence: NONE) · get_code("27659a9c234ffb11")
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 MemEIC/MemEIC, 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. A run is recorded.

check_bias_among_ls Ran
ran on a synthesized input
llm-editing/editing-attack/code/harm_util.py
code served (permissive licence) · get_code("fcf9ddd243a0bbd0")
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 torch

torch.manual_seed(0)

# The function check_bias_among_ls takes a single parameter 'text' which should be a string
# It checks if any element from ls_unbiased_ans appears in the lowercased text

# Test case 1: text containing a biased answer
text_with_bias = "The answer is unknown to me"

ARGS = {
    "text": text_with_bias
}
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 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
evaluate_response Ran
ran on a synthesized input
llm-editing/editing-attack/code/editor_new_eval.py
code served (permissive licence) · get_code("59e06d75c10101da")
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 torch
import torch.nn as nn

torch.manual_seed(0)

# Stand-in tokenizer class
class MockTokenizer:
    def __init__(self):
        self.eos_token_id = 2
    
    def convert_tokens_to_ids(self, token):
        return 2
    
    def apply_chat_template(self, messages, add_generation_prompt=True, return_tensors=None):
        # Return a small tensor of token ids
        tensor = torch.tensor([[1, 2, 3, 4, 5]], dtype=torch.long)
        return tensor
    
    def decode(self, token_ids, skip_special_tokens=False):
        return "1"

# Stand-in model class
class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def generate(self, input_ids, max_new_tokens=1, eos_token_id=None, do_sample=False, pad_token_id=None):
        # Return token ids that include the input plus generated tokens
        batch_size = input_ids.shape[0]
        generated = torch.cat([input_ids, torch.full((batch_size, max_new_tokens), 1, dtype=torch.long)], dim=1)
        return generated

# Stand-in hparams
hparams = {
    "learning_rate": 1e-3,
    "batch_size": 2
}

# Create mock model and tokenizer
model_eval = MockModel()
tok_eval = MockTokenizer()

# Test inputs
prompt_qa = "What is the capital of France?"
output_qa = "Paris"
label = "paris"
device_eval = torch.device("cpu")

ARGS = {
    "hparams": hparams,
    "model_eval": model_eval,
    "tok_eval": tok_eval,
    "prompt_qa": prompt_qa,
    "output_qa": output_qa,
    "label": label,
    "device_eval": device_eval
}
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 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
get_response Ran
ran on a synthesized input
llm-editing/editing-attack/code/editor_new_eval.py
code served (permissive licence) · get_code("446fc0f34e412c18")
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 torch.nn as nn

torch.manual_seed(0)

class MockTokenizer:
    def __init__(self):
        self.eos_token_id = 2
    
    def apply_chat_template(self, messages, add_generation_prompt=False, return_tensors=None):
        # Return a small tensor of token ids
        tokens = torch.tensor([[1, 5, 7, 2]], dtype=torch.long)
        return tokens
    
    def convert_tokens_to_ids(self, token):
        # Return a mock id for the eot token
        return 3
    
    def decode(self, token_ids, skip_special_tokens=False):
        # Return a mock decoded string
        return "mock response"

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def generate(self, input_ids, max_new_tokens=1, eos_token_id=None, do_sample=False, pad_token_id=None):
        # Return mock generated token ids
        # input_ids shape: (1, seq_len)
        batch_size = input_ids.shape[0]
        seq_len = input_ids.shape[1]
        # Generate output with original tokens plus new tokens
        new_tokens = torch.full((batch_size, seq_len + max_new_tokens), 5, dtype=torch.long)
        return new_tokens

class MockHParams:
    def __init__(self):
        self.hidden_size = 4
        self.num_layers = 1

messages = [
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi"}
]

hparams = MockHParams()
model = MockModel()
tok = MockTokenizer()
device = "cpu"

ARGS = {
    "hparams": hparams,
    "model": model,
    "tok": tok,
    "messages": messages,
    "device": device,
    "max_new_tokens": 1
}
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
hierarchical_subsequence Ran
ran on a synthesized input
llm-editing/editing-attack/code/easyeditor/util/nethook.py
code served (permissive licence) · get_code("920b394e7ad80c53")
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 Sequential. 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 torch.nn as nn
from collections import OrderedDict

torch.manual_seed(0)

# Create a nested Sequential model to test hierarchical_subsequence
layer1 = nn.Linear(4, 3)
layer2 = nn.Linear(3, 2)
layer3 = nn.Linear(2, 5)

nested_seq = nn.Sequential(
    OrderedDict([
        ('linear1', layer1),
        ('linear2', layer2),
        ('linear3', layer3)
    ])
)

# Test case 1: No parameters (should return full sequential)
ARGS = {
    'sequential': nested_seq,
    'first': None,
    'last': None,
    'after': None,
    'upto': None,
    'share_weights': False,
    'depth': 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 Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 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
kl_loc_loss Ran
ran; our contract draft was wrong, not the code
llm-editing/editing-attack/code/easyeditor/trainer/losses.py
code served (permissive licence) · get_code("bcc592aa5c1e2bb3")
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 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)

# Test case 1: Non-sequence (2D), single output dimension
pre_2d = torch.randn(2, 1)
post_2d = torch.randn(2, 1)

ARGS = {
    "pre": pre_2d,
    "post": post_2d,
    "mask": None,
}
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 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
load_api_key Ran
ran on a synthesized input
llm-editing/editing-attack/code/editor_new_eval.py
code served (permissive licence) · get_code("023650f8cf944654")
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 json
import tempfile
import os

# Create a temporary directory and a mock API key file
temp_dir = tempfile.mkdtemp()
api_key_file = os.path.join(temp_dir, 'api_key.json')

# Create the API key JSON file with test data
api_data = {
    'openai_key': 'sk-test-key-12345',
    'github_token': 'ghp_test_token_67890',
    'other_key': 'some_value'
}

with open(api_key_file, 'w') as f:
    json.dump(api_data, f)

ARGS = {
    'key': 'openai_key',
    'file_path': api_key_file
}
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
masked_mean Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
llm-editing/editing-attack/code/easyeditor/trainer/losses.py
code served (permissive licence) · get_code("e078eeed20f03838")
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 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

torch.manual_seed(0)

# Create small tensors for the function
values = torch.randn(2, 4)
mask = torch.randint(0, 2, (2, 4)).bool()

ARGS = {
    "values": values,
    "mask": mask,
}
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 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
recursive_copy Ran
ran; our contract draft was wrong, not the code
llm-editing/editing-attack/code/easyeditor/util/nethook.py
pointer only (licence: MIT); the stored copy get_code serves, from TrustedLLM/UnKE, is pointer only (licence: NONE) · get_code("70f6ab8bde55420e")
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 TrustedLLM/UnKE, 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. A run is recorded.

subsequence Ran
ran; our contract draft was wrong, not the code
llm-editing/editing-attack/code/easyeditor/util/nethook.py
code served (permissive licence) · get_code("440ff98c2b1ae1aa")
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 Sequential. 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 torch.nn as nn
import copy
from collections import OrderedDict

torch.manual_seed(0)

# Build a nested Sequential model with multiple layers
sequential = nn.Sequential(
    OrderedDict([
        ('layer1', nn.Linear(4, 3)),
        ('layer2', nn.ReLU()),
        ('layer3', nn.Linear(3, 2)),
        ('nested', nn.Sequential(
            OrderedDict([
                ('sublayer1', nn.Linear(2, 2)),
                ('sublayer2', nn.ReLU()),
            ])
        )),
        ('layer4', nn.Linear(2, 1)),
    ])
)

ARGS = {
    'sequential': sequential,
    'first_layer': None,
    'last_layer': None,
    'after_layer': None,
    'upto_layer': None,
    'single_layer': None,
    'share_weights': False,
}
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 Sequential — a Sequential output has no array shape of its own; none is 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_model Not yet run
unverified (harvested, no recorded run)
llm-editing/editing-attack/code/easyeditor/trainer/models.py
pointer only (licence: MIT); the stored copy get_code serves, from llm-editing/HalluEditBench, is pointer only (licence: NONE) · get_code("d2e4c1979dd70b9a")
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 llm-editing/HalluEditBench, 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.

get_tokenizer Not yet run
unverified (harvested, no recorded run)
llm-editing/editing-attack/code/easyeditor/trainer/models.py
code served (permissive licence) · get_code("2defe42cf1a7f301")
How it ran

no invocation recorded.

qa_acc_eval_llm Not yet run
unverified (harvested, no recorded run)
llm-editing/editing-attack/code/harm_eval_natural_questions.py
code served (permissive licence) · get_code("2ff23c4b88f49f66")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 8 are also present in other papers whose code we hold (8 with at least one paper dated earlier, 8 the same month, 8 later) and 5 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
subsequence 49 other papers
25 dated earlier, 1 the same month, 22 later, 1 same year, order not determinable
2007.15646 (2020-07)
Rewriting a Deep Generative Model
2009.05041 (2020-09)
Understanding the Role of Individual Units in a Deep Neural Network
2110.04292 (2021-10)
Toward a Visual Concept Vocabulary for GAN Latent Space
and 46 more not named here
recursive_copy 48 other papers
24 dated earlier, 1 the same month, 22 later, 1 same year, order not determinable
2207.02774 (2022-07)
Local Relighting of Real Scenes
2210.07229 (2022-10)
Mass-Editing Memory in a Transformer
2302.12461 (2023-02)
Analyzing And Editing Inner Mechanisms Of Backdoored Language Models
and 45 more not named here
hierarchical_subsequence 46 other papers
22 dated earlier, 1 the same month, 22 later, 1 same year, order not determinable
2210.07229 (2022-10)
Mass-Editing Memory in a Transformer
2302.12461 (2023-02)
Analyzing And Editing Inner Mechanisms Of Backdoored Language Models
2311.09053 (2023-11)
Assessing Knowledge Editing in Language Models via Relation Perspective
and 43 more not named here
binary_log_probs 28 other papers
14 dated earlier, 1 the same month, 12 later, 1 same year, order not determinable
2210.07229 (2022-10)
Mass-Editing Memory in a Transformer
2310.10322 (2023-10)
Evaluating the Reversal Curse in Model Editing
2311.09053 (2023-11)
Assessing Knowledge Editing in Language Models via Relation Perspective
and 25 more not named here
kl_loc_loss 16 other papers
8 dated earlier, 1 the same month, 7 later
2310.10322 (2023-10)
Evaluating the Reversal Curse in Model Editing
2401.04700 (2024-01)
Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue
2401.17623 (2024-01)
Neighboring Perturbations of Knowledge Editing on Large Language Models
and 13 more not named here
get_tokenizer 14 other papers
8 dated earlier, 1 the same month, 5 later
2310.10322 (2023-10)
Evaluating the Reversal Curse in Model Editing
2401.04700 (2024-01)
Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue
2401.17623 (2024-01)
Neighboring Perturbations of Knowledge Editing on Large Language Models
and 11 more not named here
masked_mean 10 other papers
3 dated earlier, 1 the same month, 6 later
2206.06520 (2022-06)
Memory-Based Model Editing at Scale
2310.10322 (2023-10)
Evaluating the Reversal Curse in Model Editing
2406.11194 (2024-06)
In-Context Editing: Learning Knowledge from Self-Induced Distributions
and 7 more not named here
get_model 8 other papers
5 dated earlier, 1 the same month, 2 later
2310.10322 (2023-10)
Evaluating the Reversal Curse in Model Editing
2401.04700 (2024-01)
Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue
2402.11905 (2024-02)
Learning to Edit: Aligning LLMs with Knowledge Editing
and 5 more not named here

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

Large Language Models (LLMs) have emerged as a new information channel. Meanwhile, one critical but under-explored question is: Is it possible to bypass the safety alignment and inject harmful information into LLMs stealthily? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the first risk, we find that editing attacks can inject both commonsense and long-tail misinformation into LLMs, and the effectiveness for the former one is particularly high. For the second risk, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can degrade the overall fairness. Then, we further illustrate the high stealthiness of editing attacks. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs and the feasibility of disseminating misinformation or bias with LLMs as new channels.

For agents

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

get_harvested_code_for_paper("2407.20224")
get_code_for_paper("2407.20224")
have("2407.20224")

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

curl https://syntology.ai/api/ran/2407.20224.json

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

[![Syntology run record](https://syntology.ai/api/ran/2407.20224.svg)](https://syntology.ai/paper/2407.20224)

Connect an agent — have() is free.