SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.12935 · 2024

ChatBug: A Common Vulnerability of Aligned LLMs Induced by Chat Templates

arXiv · PDF · Open in the Atlas

Code that ran

3 of 5 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, 1 violated it and 0 have no contract verdict; 2 failures of our instrument; 2 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2406.12935.json · badge: /api/ran/2406.12935.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
uw-nsl/ChatBug canonical (classifier label)
classifier basis: census
3 ran · 2 no recorded run
FunctionStatusWhere it lives
is_cuda_OOM Ran
ran, violated the contract we drafted
uw-nsl/ChatBug/utils.py
code served (permissive licence) · get_code("54efe04544a5282a")
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

# Test case 1: Valid CUDA OOM error
cuda_oom_error = RuntimeError('CUDA out of memory')

# Test case 2: RuntimeError without CUDA OOM message
other_runtime_error = RuntimeError('some other error')

# Test case 3: RuntimeError with multiple args (should return False)
multi_arg_error = RuntimeError('error1', 'error2')

# Test case 4: Non-RuntimeError exception
type_error = TypeError('not a runtime error')

# We need to test the function with different error types
# The function checks: isinstance(e, RuntimeError) and len(e.args) == 1 and ('CUDA out of memory' in e.args[0])

ARGS = {
    'e': cuda_oom_error
}
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
load_jinja_template Ran
ran; our contract draft was wrong, not the code
uw-nsl/ChatBug/utils.py
code served (permissive licence) · get_code("63d4b0b8df6ed4bd")
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 Template. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from jinja2.sandbox import ImmutableSandboxedEnvironment

# The function load_jinja_template takes a template_str parameter
# It creates a Jinja2 environment and loads a template from the string
# The template_str should be a valid Jinja2 template string

# Simple valid Jinja2 template examples:
simple_template = "Hello {{ name }}!"
loop_template = "{% for item in items %}{{ item }}{% endfor %}"
conditional_template = "{% if x > 5 %}big{% else %}small{% endif %}"
exception_template = "{% if error %}{{ raise_exception('Custom error') }}{% endif %}"

# Use a simple, valid template string
template_str = "Hello {{ name }}! {% if show_message %}The value is {{ value }}.{% endif %}"

ARGS = {
    "template_str": template_str
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jinja2, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Template — a Template output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Template 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
try_with_excuable_batch_size Ran
ran; our contract draft was wrong, not the code
uw-nsl/ChatBug/utils.py
code served (permissive licence) · get_code("39823c70f680a83a")
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 function. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type 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 functools import wraps

torch.manual_seed(0)

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

class MockProcessor:
    def __init__(self):
        self.batch_size = 4
        self.model = MockModel()
    
    @staticmethod
    def try_with_excuable_batch_size(method_func):
        @wraps(method_func)
        def __impl(self, *args, **kwargs):
            while self.batch_size >= 1:
                try:
                    return method_func(self, *args, **kwargs)
                except Exception as e:
                    def is_cuda_OOM(e):
                        return isinstance(e, RuntimeError) and len(e.args) == 1 and ('CUDA out of memory' in e.args[0])
                    
                    if is_cuda_OOM(e):
                        if self.batch_size == 1:
                            print('OOM error, failed with batch size 1')
                            raise e
                        else:
                            print('OOM error, retrying with smaller batch size')
                            self.batch_size = self.batch_size // 2
                            print(f'New batch size: {self.batch_size}')
                            torch.cuda.empty_cache()
                    else:
                        print(f'Error: {e}')
                        raise e
        return __impl

processor = MockProcessor()

x_batch = torch.randn(2, 4)
y_batch = torch.randn(2, 3)

ARGS = {
    'method_func': lambda self, x, y: self.model(x),
}
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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
mask_word_stop Not yet run
unverified (harvested, no recorded run)
uw-nsl/ChatBug/artprompt.py
code served (permissive licence) · get_code("2abb20aa39df0144")
How it ran

no invocation recorded.

override_config Not yet run
unverified (harvested, no recorded run)
uw-nsl/ChatBug/chatbug.py
code served (permissive licence) · get_code("7f2c5700d03f738a")
How it ran

no invocation recorded.

Also present in other papers' code

Of 5 samples lifted from this paper's linked repositories, 5 carry a function identity and 0 do not. Of the 5 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 4 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
mask_word_stop 1 other paper
1 dated earlier
2402.11753 (2024-02)
ArtPrompt: ASCII Art-based Jailbreak Attacks against Aligned LLMs WARNING: This paper contains model outputs that may be considered offensive

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) are expected to follow instructions from users and engage in conversations. Techniques to enhance LLMs' instruction-following capabilities typically fine-tune them using data structured according to a predefined chat template. Although chat templates are shown to be effective in optimizing LLM performance, their impact on safety alignment of LLMs has been less understood, which is crucial for deploying LLMs safely at scale. In this paper, we investigate how chat templates affect safety alignment of LLMs. We identify a common vulnerability, named ChatBug, that is introduced by chat templates. Our key insight to identify ChatBug is that the chat templates provide a rigid format that need to be followed by LLMs, but not by users. Hence, a malicious user may not necessarily follow the chat template when prompting LLMs. Instead, malicious users could leverage their knowledge of the chat template and accordingly craft their prompts to bypass safety alignments of LLMs. We develop two attacks to exploit the ChatBug vulnerability. We demonstrate that a malicious user can exploit the ChatBug vulnerability of eight state-of-the-art (SOTA) LLMs and effectively elicit unintended responses from these models. Moreover, we show that ChatBug can be exploited by existing jailbreak attacks to enhance their attack success rates. We investigate potential countermeasures to ChatBug. Our results show that while adversarial training effectively mitigates the ChatBug vulnerability, the victim model incurs significant performance degradation. These results highlight the trade-off between safety alignment and helpfulness. Developing new methods for instruction tuning to balance this trade-off is an open and critical direction for future research

For agents

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

get_harvested_code_for_paper("2406.12935")
get_code_for_paper("2406.12935")
have("2406.12935")

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

curl https://syntology.ai/api/ran/2406.12935.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.12935.svg)](https://syntology.ai/paper/2406.12935)

Connect an agent — have() is free.