SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2403.09792 · 2024

Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models

arXiv · PDF · Open in the Atlas

Code that ran

7 of 10 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, 2 violated it and 0 have no contract verdict; 4 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/2403.09792.json · badge: /api/ran/2403.09792.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
aoidragon/hades canonical (classifier label)
classifier basis: classifier_2025
2 ran · 1 no recorded run
copy not recorded community
not in the paper's declared repositories
3 ran · 2 no recorded run
rucaibox/hades reimplementation (classifier label)
classifier basis: classifier_2025
2 ran · 0 no recorded run
FunctionStatusWhere it lives
concatenate_images Ran
ran; our contract draft was wrong, not the code
aoidragon/hades/white_box/generate_init_image.py
code served (permissive licence) · get_code("feb9e7a6ed55eff5")
How it ran

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

from PIL import Image

# Create two simple RGB images with distinct sizes
image1 = Image.new('RGB', (100, 80), color='red')
image2 = Image.new('RGB', (120, 60), color='blue')

ARGS = {
    'image1': image1,
    'image2': image2,
}
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 Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_caps Ran
ran; our contract draft was wrong, not the code
this paper's copy was not recorded; identical code first harvested from llava-annonymous/llava
pointer only · get_code("94fda0761ce0d5b9")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

generate_white_noise_image Ran
ran, honoured the contract we drafted
aoidragon/hades/white_box/generate_init_image.py
code served (permissive licence) · get_code("cb498d178cdcc7a7")
How it ran

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

import numpy as np
from PIL import Image

# Set random seed for reproducibility
np.random.seed(0)

# Define arguments for generate_white_noise_image
ARGS = {
    "width": 64,
    "height": 48
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_pred_idx Ran
ran, violated the contract we drafted
rucaibox/hades/eval/evaluate/inference/llava/eval/eval_science_qa.py
code served (permissive licence) · get_code("36e875f85baa4565")
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 int, ndim 0, dtype family int, 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 random

def get_pred_idx(prediction, choices, options):
    """
    Get the index (e.g. 2) from the prediction (e.g. 'C')
    """
    if prediction in options[:len(choices)]:
        return options.index(prediction)
    else:
        return -1
        return random.choice(range(len(choices)))

# Setup code
prediction = 'B'
choices = ['option1', 'option2', 'option3']
options = ['A', 'B', 'C', 'D']

ARGS = {
    'prediction': prediction,
    'choices': choices,
    'options': options
}
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 int, ndim 0, dtype family int, 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
parse_score Ran
ran, violated the contract we drafted
this paper's copy was not recorded; identical code first harvested from TempleX98/MoVA
pointer only · get_code("763f6fe20bd52fbb")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

prompt_processor Ran
ran; our contract draft was wrong, not the code
rucaibox/hades/eval/evaluate/inference/llava/eval/eval_textvqa.py
code served (permissive licence) · get_code("1e68840d6bbdface")
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 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 prompt_processor(prompt):
    if prompt.startswith('OCR tokens: '):
        pattern = 'Question: (.*?) Short answer:'
        match = re.search(pattern, prompt, re.DOTALL)
        question = match.group(1)
    elif 'Reference OCR token: ' in prompt and len(prompt.split('\n')) == 3:
        if prompt.startswith('Reference OCR token:'):
            question = prompt.split('\n')[1]
        else:
            question = prompt.split('\n')[0]
    elif len(prompt.split('\n')) == 2:
        question = prompt.split('\n')[0]
    else:
        assert False
    return question.lower()


# Test case 1: prompt starting with 'OCR tokens: '
prompt_1 = 'OCR tokens: Question: What is this? Short answer: test'

# Test case 2: prompt with 'Reference OCR token: ' and 3 lines, starting with it
prompt_2 = 'Reference OCR token: \nWhat is the answer?\nMore info'

# Test case 3: prompt with exactly 2 lines
prompt_3 = 'What is the question?\nSome context'

ARGS = {
    'prompt': prompt_1
}
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 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
split_to_even_chunks Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
this paper's copy was not recorded; identical code first harvested from haotian-liu/LLaVA
pointer only · get_code("10893c4608c08075")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

generate_and_save_white_noise_image Not yet run
unverified (harvested, no recorded run)
aoidragon/hades/white_box/generate_init_image.py
code served (permissive licence) · get_code("49d1501fe6d6a841")
How it ran

no invocation recorded.

get_mm_adapter_state_maybe_zero_3 Not yet run
unverified (harvested, no recorded run)
this paper's copy was not recorded; identical code first harvested from haotian-liu/LLaVA
pointer only · get_code("bb35e3ac741bb2c9")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

maybe_zero_3 Not yet run
unverified (harvested, no recorded run)
this paper's copy was not recorded; identical code first harvested from haotian-liu/LLaVA
pointer only · get_code("735025744c1ab0cf")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 7 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 4 the same month, 7 later) and 3 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_mm_adapter_state_maybe_zero_3 43 other papers
10 dated earlier, 1 the same month, 32 later
2304.05884 (2023-04)
Unicom: Universal and Compact Representation Learning for Image Retrieval
2304.08485 (2023-04)
Visual Instruction Tuning
2310.03744 (2023-10)
Improved Baselines with Visual Instruction Tuning
and 40 more not named here
maybe_zero_3 42 other papers
10 dated earlier, 1 the same month, 31 later
2304.05884 (2023-04)
Unicom: Universal and Compact Representation Learning for Image Retrieval
2304.08485 (2023-04)
Visual Instruction Tuning
2310.03744 (2023-10)
Improved Baselines with Visual Instruction Tuning
and 39 more not named here
split_to_even_chunks 40 other papers
9 dated earlier, 1 the same month, 30 later
2304.05884 (2023-04)
Unicom: Universal and Compact Representation Learning for Image Retrieval
2304.08485 (2023-04)
Visual Instruction Tuning
2310.03744 (2023-10)
Improved Baselines with Visual Instruction Tuning
and 37 more not named here
parse_score 27 other papers
8 dated earlier, 2 the same month, 17 later
2309.09958 (2023-09)
An Empirical Study of Scaling Instruct-Tuned Large Multimodal Models
2310.03744 (2023-10)
Improved Baselines with Visual Instruction Tuning
2312.03818 (2023-12)
Alpha-CLIP: A CLIP Model Focusing on Wherever You Want
and 24 more not named here
convert_caps 3 other papers
1 dated earlier, 2 later
2304.08485 (2023-04)
Visual Instruction Tuning
2412.00876 (2024-12)
Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification
2503.15621 (2025-03)
LLaVA-MORE: A Comparative Study of LLMs and Visual Backbones for Enhanced Visual Instruction Tuning
get_pred_idx 3 other papers
1 dated earlier, 2 later
2304.08485 (2023-04)
Visual Instruction Tuning
2412.00876 (2024-12)
Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification
2503.15621 (2025-03)
LLaVA-MORE: A Comparative Study of LLMs and Visual Backbones for Enhanced Visual Instruction Tuning
prompt_processor 1 other paper
1 later
2412.00876 (2024-12)
Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification

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

In this paper, we study the harmlessness alignment problem of multimodal large language models (MLLMs). We conduct a systematic empirical analysis of the harmlessness performance of representative MLLMs and reveal that the image input poses the alignment vulnerability of MLLMs. Inspired by this, we propose a novel jailbreak method named HADES, which hides and amplifies the harmfulness of the malicious intent within the text input, using meticulously crafted images. Experimental results show that HADES can effectively jailbreak existing MLLMs, which achieves an average Attack Success Rate (ASR) of 90.26% for LLaVA-1.5 and 71.60% for Gemini Pro Vision. Our code and data are available at https://github.com/RUCAIBox/HADES.

For agents

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

get_harvested_code_for_paper("2403.09792")
get_code_for_paper("2403.09792")
have("2403.09792")

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

curl https://syntology.ai/api/ran/2403.09792.json

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

[![Syntology run record](https://syntology.ai/api/ran/2403.09792.svg)](https://syntology.ai/paper/2403.09792)

Connect an agent — have() is free.