SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2408.08926 · ICLR · 2025

Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models

Percy Liang, Daniel Ho, Dan Boneh, Andy Zhang, Neil Perry, Riya Dulepet, Joey Ji, Celeste Menders, Justin Lin, Eliot Jones, Gashon Hussein, Samantha Liu, and 15 more

arXiv · PDF · Open in the Atlas

Code that ran

10 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; 4 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 6 failures of our instrument; 0 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/2408.08926.json · badge: /api/ran/2408.08926.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
andyzorigin/cyber-bench canonical (classifier label)
classifier basis: census
9 ran · 0 no recorded run
andyzorigin/cybench canonical (classifier label)
classifier basis: classifier_2025
1 ran · 0 no recorded run
FunctionStatusWhere it lives
analyze_challenge_completion Ran
ran; our contract draft was wrong, not the code
andyzorigin/cyber-bench/analytics/analyze_benchmark_run.py
code served (permissive licence) · get_code("a7e69c0c5022bb4f")
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 torch
import numpy as np

torch.manual_seed(0)

ARGS = {
    'data': {
        'challenge': {
            'subtasks': [
                {
                    'question': 'What is 2+2?',
                    'answer': '4'
                },
                {
                    'question': 'What is 3+5?',
                    'answer': '8'
                },
                {
                    'question': 'What is the flag?',
                    'answer': 'FLAG{correct}'
                }
            ],
            'categories': ['math', 'logic', 'puzzle'],
            'difficulty': 'medium'
        },
        'subtask_completions': {
            'subtask_1': {
                'completion_status': 'answered',
                'solved_with_hint': False,
                'time_taken_in_ms': 1500,
                'iterations': [
                    {
                        'execution_output': {
                            'stdout': '4'
                        }
                    }
                ]
            },
            'subtask_2': {
                'completion_status': 'answered',
                'solved_with_hint': True,
                'time_taken_in_ms': 2000,
                'iterations': [
                    {
                        'execution_output': {
                            'stdout': '8'
                        }
                    }
                ]
            },
            'subtask_3': {
                'completion_status': 'answered',
                'solved_with_hint': False,
                'time_taken_in_ms': 3000,
                'iterations': [
                    {
                        'execution_output': {
                            'stdout': 'FLAG{correct}'
                        }
                    }
                ]
            }
        },
        'usage': {
            'total_tokens': 1250
        },
        'time_taken_in_ms': 6500
    }
}
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 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
analyze_completion Ran
ran; our contract draft was wrong, not the code
andyzorigin/cyber-bench/analytics/analyze_benchmark_run.py
code served (permissive licence) · get_code("4fe2313caf422564")
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 torch
import numpy as np

# Build the arguments for analyze_completion function
# The function expects:
# - data: List[Dict] - a list of challenge completion data dictionaries
# - solve_times_dict: Dict[str, str] - a dictionary mapping challenge paths to solve times

# First, let's construct a single challenge data entry that matches what analyze_challenge_completion expects
def create_challenge_data_entry():
    return {
        'challenge': {
            'name': 'Challenge 1',
            'path': '/challenges/challenge_1',
            'difficulty': 'medium',
            'categories': ['algorithm'],
            'subtasks': [
                {
                    'question': 'What is 2+2?',
                    'answer': '4'
                },
                {
                    'question': 'What is 3*3?',
                    'answer': '9'
                }
            ]
        },
        'subtask_completions': {
            'subtask_0': {
                'completion_status': 'answered',
                'solved_with_hint': False,
                'time_taken_in_ms': 1500,
                'iterations': [
                    {
                        'execution_output': {
                            'stdout': '4'
                        }
                    }
                ]
            },
            'subtask_1': {
                'completion_status': 'answered',
                'solved_with_hint': True,
                'time_taken_in_ms': 2000,
                'iterations': [
                    {
                        'execution_output': {
                            'stdout': '9'
                        }
                    }
                ]
            }
        },
        'usage': {
            'total_tokens': 500
        },
        'time_taken_in_ms': 3500
    }

def create_challenge_data_entry_2():
    return {
        'challenge': {
            'name': 'Challenge 2',
            'path': '/challenges/challenge_2',
            'difficulty': 'hard',
            'categories': ['string'],
            'subtasks': [
                {
                    'question': 'Reverse the string',
                    'answer': 'olleh'
                }
            ]
        },
        'subtask_completions': {
            'subtask_0': {
                'completion_status': 'answered',
                'solved_with_hint': False,
                'time_taken_in_ms': 1000,
                'iterations': [
                    {
                        'execution_output': {
                            'stdout': 'olleh'
                        }
                    }
                ]
            }
        },
        'usage': {
            'total_tokens': 300
        },
        'time_taken_in_ms': 1000
    }

# Construct the arguments
data = [create_challenge_data_entry(), create_challenge_data_entry_2()]

solve_times_dict = {
    '/challenges/challenge_1': '2023-01-01 10:00:00',
    '/challenges/challenge_2': '2023-01-01 11:30:00'
}

ARGS = {
    'data': data,
    'solve_times_dict': solve_times_dict
}
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 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
calculate_percentage Ran
ran, honoured the contract we drafted
andyzorigin/cyber-bench/grade_benchmark.py
code served (permissive licence) · get_code("7b2dfc15a8b7acab")
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 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.

import torch

torch.manual_seed(0)

# correct_flags should be a list-like object where len() works
correct_flags = [True, True, False, True]

# total_flags should be a number
total_flags = 4

ARGS = {
    "correct_flags": correct_flags,
    "total_flags": total_flags
}
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 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
duration_to_seconds Ran
ran, honoured the contract we drafted
andyzorigin/cyber-bench/grading/guided_run.py
code served (permissive licence) · get_code("9bcfded361d2b865")
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 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 torch

def duration_to_seconds(duration):
    parts = duration.split(':')
    parts = [int(part) for part in parts]
    if len(parts) == 2:
        return parts[0] * 60 + parts[1]
    elif len(parts) == 3:
        return parts[0] * 3600 + parts[1] * 60 + parts[2]
    else:
        raise ValueError('Invalid duration format')

ARGS = {
    'duration': '1:30'
}
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 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
duration_to_seconds Ran
ran, honoured the contract we drafted
andyzorigin/cyber-bench/grading/unguided_run.py
code served (permissive licence) · get_code("7fb8980bde1da170")
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 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 torch

def duration_to_seconds(duration):
    parts = duration.split(':')
    parts = [int(part) for part in parts]
    if len(parts) == 2:
        return parts[0] * 60 + parts[1]
    elif len(parts) == 3:
        return parts[0] * 3600 + parts[1] * 60 + parts[2]
    else:
        raise ValueError('Invalid duration format')
        import pandas as pd

ARGS = {
    'duration': '1:30'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas, torch — 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
load_data Ran
ran; our contract draft was wrong, not the code
andyzorigin/cyber-bench/grade_benchmark.py
code served (permissive licence) · get_code("3fd3e80df53a806b")
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 os
import json
import tempfile
import shutil
from pathlib import Path

# Create a temporary directory structure with test JSON files
temp_dir = tempfile.mkdtemp()

# Create test data structure matching what load_data expects
test_data_1 = {
    'challenge_run_input': {
        'agent_config': {
            'deployment_name': 'gpt-4'
        }
    },
    'challenge': {
        'path': 'challenge_1',
        'subtasks': [
            {'answer': '{flag_1}'},
            {'answer': '{flag_2}'}
        ]
    },
    'usage': {
        'input_tokens_used': 1000,
        'output_tokens_used': 500
    },
    'subtask_completions': {
        'subtask_0': {
            'iterations': [
                {
                    'model_input': {'num_tokens': 3000},
                    'model_response': {'num_tokens': 400, 'value': 'response_1'}
                },
                {
                    'model_input': {'num_tokens': 2000},
                    'model_response': {'num_tokens': 300, 'value': 'response_2'}
                }
            ],
            'score': 1
        },
        'subtask_1': {
            'iterations': [
                {
                    'model_input': {'num_tokens': 2500},
                    'model_response': {'num_tokens': 350, 'value': 'response_3'}
                }
            ],
            'score': 0
        }
    }
}

test_data_2 = {
    'task_run_input': {
        'agent_config': {
            'deployment_name': 'claude-3'
        }
    },
    'task': {
        'path': 'challenge_2',
        'subtasks': [
            {'answer': 'flag_without_braces'},
            {'answer': '{flag_x}'}
        ]
    },
    'usage': {
        'input_tokens_used': 800,
        'output_tokens_used': 600
    },
    'subtask_completions': {
        'subtask_0': {
            'iterations': [
                {
                    'model_input': {'num_tokens': 1500},
                    'model_response': {'num_tokens': 200, 'value': 'response_a'}
                }
            ],
            'score': 1
        },
        'subtask_1': {
            'iterations': [
                {
                    'model_input': {'num_tokens': 2000},
                    'model_response': {'num_tokens': 250, 'value': 'response_b'}
                }
            ],
            'score': 1
        }
    }
}

test_data_3 = {
    'challenge_run_input': {
        'agent_config': {
            'deployment_name': 'google-gemini-1.5-pro'
        }
    },
    'challenge': {
        'path': 'challenge_3',
        'subtasks': [
            {'answer': '{result}'}
        ]
    },
    'usage': {
        'input_tokens_used': 1200,
        'output_tokens_used': 800
    },
    'subtask_completions': {
        'subtask_0': {
            'iterations': [
                {
                    'model_input': {'num_tokens': 6500},
                    'model_response': {'num_tokens': 2500, 'value': 'result_output'},
                    'execution_output': {
                        'stdout': 'result'
                    }
                }
            ],
            'score': 1
        }
    }
}

# Write test JSON files
subdir = os.path.join(temp_dir, 'data')
os.makedirs(subdir, exist_ok=True)

with open(os.path.join(subdir, 'test_1.json'), 'w') as f:
    json.dump(test_data_1, f)

with open(os.path.join(subdir, 'test_2.json'), 'w') as f:
    json.dump(test_data_2, f)

with open(os.path.join(subdir, 'test_3.json'), 'w') as f:
    json.dump(test_data_3, f)

ARGS = {
    'directory_name': 'data'
}
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 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
load_list_from_file Ran
ran; our contract draft was wrong, not the code
andyzorigin/cybench/run_benchmark.py
code served (permissive licence) · get_code("80047d3cba4ac7d9")
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 tempfile
import os

# Create a temporary file with some test content
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt')
temp_file.write("line one\n")
temp_file.write("line two\n")
temp_file.write("  \n")
temp_file.write("line three\n")
temp_file.close()

ARGS = {
    'filename': temp_file.name
}
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
read_fastest_solve_times Ran
ran; our contract draft was wrong, not the code
andyzorigin/cyber-bench/analytics/analyze_benchmark_run.py
code served (permissive licence) · get_code("433972f6b994dc8e")
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 json
import tempfile
import os

# Create a temporary JSON file with valid data structure
temp_dir = tempfile.mkdtemp()
json_file = os.path.join(temp_dir, "fastest_times.json")

# Create sample data matching the expected structure
fastest_times_data = [
    {
        "challenge_dir": "challenge_1",
        "challenge_fastest_solve_time": 42.5
    },
    {
        "challenge_dir": "challenge_2",
        "challenge_fastest_solve_time": 35.2
    },
    {
        "challenge_dir": "challenge_3",
        "challenge_fastest_solve_time": 50.1
    }
]

# Write the JSON file
with open(json_file, 'w') as f:
    json.dump(fastest_times_data, f)

ARGS = {
    "json_file": json_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 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
read_metadata Ran
ran; our contract draft was wrong, not the code
andyzorigin/cyber-bench/run_task.py
code served (permissive licence) · get_code("313007004d0ed238")
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 json
import logging
import os
import sys
import tempfile
import shutil

logging.basicConfig(level=logging.INFO)

# Create a temporary directory with the required metadata structure
temp_dir = tempfile.mkdtemp()
metadata_dir = os.path.join(temp_dir, 'metadata')
os.makedirs(metadata_dir, exist_ok=True)

# Create a valid metadata.json file
metadata_content = {
    "task_name": "example_task",
    "version": "1.0",
    "description": "An example metadata file",
    "parameters": {
        "param1": "value1",
        "param2": "value2"
    }
}

metadata_file_path = os.path.join(metadata_dir, 'metadata.json')
with open(metadata_file_path, 'w') as f:
    json.dump(metadata_content, f)

# Define ARGS with the task_dir parameter
ARGS = {
    'task_dir': temp_dir
}
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
time_str_to_minutes Ran
ran, honoured the contract we drafted
andyzorigin/cyber-bench/grade_benchmark.py
code served (permissive licence) · get_code("269b537229622374")
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 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.

import torch
import numpy as np

# The function time_str_to_minutes takes a single parameter: time_str
# It expects a string in the format "HH:MM:SS" (3 parts separated by colons)
# The function splits the string, converts parts to integers, and calculates total minutes

# Valid test case: 1 hour, 30 minutes, 45 seconds
# Expected: 1*60 + 30 + 45/60.0 = 60 + 30 + 0.75 = 90.75 minutes
time_str_value = "1:30:45"

ARGS = {
    "time_str": time_str_value
}
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 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

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

Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause realworld impact. Policymakers, model providers, and researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. 1 We include 40 professionallevel Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structured bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. See Author Contributions for detailed author contributions. 1 All code and data are publicly available at https://cybench.github.io.

For agents

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

get_harvested_code_for_paper("2408.08926")
get_code_for_paper("2408.08926")
have("2408.08926")

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

curl https://syntology.ai/api/ran/2408.08926.json

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

[![Syntology run record](https://syntology.ai/api/ran/2408.08926.svg)](https://syntology.ai/paper/2408.08926)

Connect an agent — have() is free.