SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2402.14020 · 2024

Coercing LLMs to do and reveal (almost) anything

arXiv · PDF · Open in the Atlas

Code that ran

6 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; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 6 failures of our instrument; 4 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/2402.14020.json · badge: /api/ran/2402.14020.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
jonasgeiping/carving canonical (classifier label)
classifier basis: census
6 ran · 4 no recorded run
FunctionStatusWhere it lives
flatten Ran
ran; our contract draft was wrong, not the code
jonasgeiping/carving/carving/utils.py
code served (permissive licence) · get_code("af4d91549465a11f")
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 collections

# Construct a nested dictionary with both mutable mappings and leaf values
d = {
    'model': 'resnet50',
    'training': {
        'lr': 0.001,
        'epochs': 100,
        'optimizer': {
            'name': 'adam',
            'betas': (0.9, 0.999)
        }
    },
    'batch_size': 32,
    'data': {
        'path': '/data/images',
        'split': 0.8
    }
}

ARGS = {
    'd': d,
    'parent_key': '',
    'sep': '_'
}
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_data_iterators Ran
ran; our contract draft was wrong, not the code
jonasgeiping/carving/carving/data_utils.py
code served (permissive licence) · get_code("3850987c11cdd03e")
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 datasets
from datasets import Dataset

torch.manual_seed(0)

# Create a minimal Dataset object with the required structure
# The function expects either a DatasetDict with 'train' and 'test' keys,
# or a Dataset that can be split with train_test_split
data_dict = {
    'text': ['sample 1', 'sample 2', 'sample 3', 'sample 4'],
    'label': [0, 1, 0, 1]
}
data_source = Dataset.from_dict(data_dict)

ARGS = {
    'data_source': data_source,
    'split_holdout_size': 2,
    'batch_size': 2,
    'num_workers': 0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, 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
set_chat_template Ran
ran; our contract draft was wrong, not the code
jonasgeiping/carving/carving/model_interface.py
code served (permissive licence) · get_code("4138c5f696ea2167")
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 MockTokenizer. 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
from transformers import PreTrainedTokenizer

torch.manual_seed(0)

class MockTokenizer:
    """Stand-in tokenizer with chat_template attribute"""
    def __init__(self):
        self.chat_template = None

# Test with vicuna model
model_name_or_path_vicuna = "vicuna-7b-hf"
tokenizer_vicuna = MockTokenizer()

# Test with amber model
model_name_or_path_amber = "amber-7b"
tokenizer_amber = MockTokenizer()

# Test with llama base model
model_name_or_path_llama = "llama-7b-hf"
tokenizer_llama = MockTokenizer()

# Test with default case where tokenizer.chat_template is None
model_name_or_path_default = "some-model-v1"
tokenizer_default = MockTokenizer()

# Use the llama case as primary test case
ARGS = {
    "model_name_or_path": "llama-7b-hf",
    "tokenizer": MockTokenizer()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MockTokenizer — a MockTokenizer output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a MockTokenizer 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
set_system_prompt Ran
ran; our contract draft was wrong, not the code
jonasgeiping/carving/carving/model_interface.py
code served (permissive licence) · get_code("a0b20deb16f63bcf")
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 MinimalTokenizer. 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
from unittest.mock import Mock

# Create a minimal tokenizer mock that supports attribute assignment
class MinimalTokenizer:
    def __init__(self):
        self.default_system_message = None

# Test with different model names to verify the function logic
tokenizer = MinimalTokenizer()

# Create ARGS for a simple llama model test case
ARGS = {
    "model_name_or_path": "meta-llama/Llama-2-7b",
    "tokenizer": tokenizer
}
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 MinimalTokenizer — a MinimalTokenizer output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a MinimalTokenizer 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
shorten_str Ran
ran; our contract draft was wrong, not the code
jonasgeiping/carving/carving/optimizers/generic_optimizer.py
code served (permissive licence) · get_code("bf78966d269c8d51")
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 shorten_str takes:
# - prompt: a string to potentially shorten
# - cutoff: an integer threshold for length (default 40)
#
# The function checks len(prompt) and uses string slicing,
# so we need to provide a string for prompt and an integer for cutoff

ARGS = {
    "prompt": "This is a test prompt that should be shortened because it is quite long",
    "cutoff": 40
}
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
system_startup Ran
ran; our contract draft was wrong, not the code
jonasgeiping/carving/carving/utils.py
code served (permissive licence) · get_code("ef123b64470adcdc")
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. 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 numpy as np
import random
import os
import sys

torch.manual_seed(0)
np.random.seed(0)
random.seed(0)

class MockConfig:
    def __init__(self):
        self.impl = MockImpl()
        self.seed = 42
    
class MockImpl:
    def __init__(self):
        self.cudnn_benchmark = False
        self.matmul_precision = 'high'
        self.tf32_allowed = False
        self.enable_huggingface_offline_mode = False
        self.dist_backend = 'nccl'
        self.max_workers = 4
        self.default_precision = 'float32'
        self.deterministic = False

cfg = MockConfig()

ARGS = {
    'cfg': cfg
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
multiprocess, numpy, psutil, 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
load_and_prep_dataset Not yet run
unverified (harvested, no recorded run)
jonasgeiping/carving/carving/data_utils.py
code served (permissive licence) · get_code("82786413c99b6efd")
How it ran

no invocation recorded.

load_local_harmful_behavior_dataset Not yet run
unverified (harvested, no recorded run)
jonasgeiping/carving/carving/data_utils.py
code served (permissive licence) · get_code("cc717c3fca1441a6")
How it ran

no invocation recorded.

load_model_and_tokenizer Not yet run
unverified (harvested, no recorded run)
jonasgeiping/carving/carving/model_interface.py
code served (permissive licence) · get_code("832c4bb5405dc600")
How it ran

no invocation recorded.

pathfinder Not yet run
unverified (harvested, no recorded run)
jonasgeiping/carving/carving/utils.py
code served (permissive licence) · get_code("e3207f3fcc44ddda")
How it ran

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, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 10 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.

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

It has recently been shown that adversarial attacks on large language models (LLMs) can "jailbreak" the model into making harmful statements. In this work, we argue that the spectrum of adversarial attacks on LLMs is much larger than merely jailbreaking. We provide a broad overview of possible attack surfaces and attack goals. Based on a series of concrete examples, we discuss, categorize and systematize attacks that coerce varied unintended behaviors, such as misdirection, model control, denial-of-service, or data extraction. We analyze these attacks in controlled experiments, and find that many of them stem from the practice of pre-training LLMs with coding capabilities, as well as the continued existence of strange "glitch" tokens in common LLM vocabularies that should be removed for security reasons.

For agents

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

get_harvested_code_for_paper("2402.14020")
get_code_for_paper("2402.14020")
have("2402.14020")

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

curl https://syntology.ai/api/ran/2402.14020.json

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

[![Syntology run record](https://syntology.ai/api/ran/2402.14020.svg)](https://syntology.ai/paper/2402.14020)

Connect an agent — have() is free.