Dynamic-Risk-Assessment/llm_ctf_solve.py at main · princeton-polaris-lab/Dynamic-Risk-Assessment · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
from datetime import datetime
import socket
import subprocess
import time
import json
import os
import argparse
import shutil
import yaml
from pathlib import Path

from nyuctf.dataset import CTFDataset
from nyuctf.challenge import CTFChallenge

from llm_ctf.ctflogging import status
from llm_ctf.backends import Backend, OpenAIBackend, AnthropicBackend, VLLMBackend
from llm_ctf.formatters import Formatter
from llm_ctf.prompts.prompts import PromptManager
from llm_ctf.environment import CTFEnvironment
from llm_ctf.conversation import CTFConversation

from nyuctf.dataset import CTFDataset
from nyuctf.challenge import CTFChallenge

def main():
    parser = argparse.ArgumentParser(
        description="Use an LLM to solve a CTF challenge",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter
    )
    model_list = []
    for b in Backend.registry.values():
        model_list += b.get_models()
    model_list = list(set(model_list))

    script_dir = Path(__file__).parent.resolve()

    parser.add_argument("--challenge", required=True, help="Name of the challenge")
    parser.add_argument("--dataset", help="Dataset JSON path. Only provide if not using the NYUCTF dataset at default path")
    parser.add_argument("-s", "--split", default="development", choices=["test", "development"], help="Dataset split to select. Only used when --dataset not provided.")
    parser.add_argument("-c", "--config", help="Config file to run the experiment")

    parser.add_argument("-q", "--quiet", action="store_true", help="don't print messages to the console")
    parser.add_argument("-d", "--debug", action="store_true", help="print debug messages")
    parser.add_argument("-M", "--model", help="the model to use (default is backend-specific)", choices=model_list)
    parser.add_argument("-C", "--container-image", default="ctfenv", help="the Docker image to use for the CTF environment")
    parser.add_argument("--container-name", default="ctf_env", help="the Docker container name to use for the CTF environment")
    parser.add_argument("-N", "--network", default="ctfnet", help="the Docker network to use for the CTF environment")
    parser.add_argument("--api-key", default=None, help="API key to use when calling the model")
    parser.add_argument("--api-endpoint", default=None, help="API endpoint URL to use when calling the model")
    parser.add_argument("--backend", default="openai", choices=Backend.registry.keys(), help="model backend to use")
    parser.add_argument("--formatter", default="xml", choices=Formatter.registry.keys(), help="prompt formatter to use")
    parser.add_argument("--prompt-set", default="default", help="set of prompts to use")
    # TODO add back hints functionality
    parser.add_argument("--hints", default=[], nargs="+", help="list of hints to provide")
    parser.add_argument("--disable-markdown", default=False, action="store_true", help="don't render Markdown formatting in messages")
    parser.add_argument("-m", "--max-rounds", type=int, default=10, help="maximum number of rounds to run")
    parser.add_argument("--max-cost", type=float, default=10, help="maximum cost of the conversation to run")
    parser.add_argument("--temperature", type=float, default=0.6, help="temperature for sampling")

    # Log directory options
    parser.add_argument("--skip-exist", action="store_true", help="Skip existing logs and experiments")
    parser.add_argument("-L", "--logdir", default=str(script_dir / "logs"), help="log directory to write the log")
    parser.add_argument("-n", "--name", help="Experiment name (creates subdir in logdir)")
    parser.add_argument("-i", "--index", help="Round index of the experiment (creates subdir in logdir)")

    args = parser.parse_args()
    config = None
    if args.config:
        try:
            with open(args.config, "r") as c:
                config = yaml.safe_load(c)
        except FileNotFoundError:
            pass

    if config:
        config_parameter = config.get("parameter", {})
        config_experiment = config.get("experiment", {})
        config_demostration = config.get("demostration", {})

        if not args.max_rounds:
            args.max_rounds = config_parameter.get("max_rounds", args.max_rounds)
        args.backend = config_parameter.get("backend", args.backend)
        if not args.model:
            args.model = config_parameter.get("model", None)
        # args.model = config_parameter.get("model", args.model)
        args.max_cost = config_parameter.get("max_cost", args.max_cost)
        if not args.name:
            args.name = config_experiment.get("name", None)
        # args.name = config_experiment.get("name", args.name)
        args.debug = config_experiment.get("debug", args.debug)
        args.skip_exist = config_experiment.get("skip_exist", args.skip_exist)
        args.hints = config_demostration.get("hints", [])

    print(args)

    status.set(quiet=args.quiet, debug=args.debug, disable_markdown=args.disable_markdown)

    if args.dataset is not None:
        dataset = CTFDataset(dataset_json=args.dataset)
    else:
        dataset = CTFDataset(split=args.split)
    challenge = CTFChallenge(dataset.get(args.challenge), dataset.basedir)

    logdir = Path(args.logdir).expanduser().resolve()
    logsubdir = []
    if args.name:
        logsubdir.append(args.name)
    if args.index:
        logsubdir.append(f"round{args.index}")
    if len(logsubdir) > 0:
        logdir = logdir / ("_".join(logsubdir))
    logdir.mkdir(parents=True, exist_ok=True)
    logfile = logdir / f"{challenge.canonical_name}.json"

    if logfile.exists() and args.skip_exist:
        status.print(f"[red bold]Challenge log {logfile} exists; skipping[/red bold]", markup=True)
        exit()

    print(f"Using model: {args.model}")
    print(f"Output path: {logfile}")
    print(f"Max rounds: {args.max_rounds}")

    environment = CTFEnvironment(challenge, args.container_image, args.network, args.container_name)
    prompt_manager = PromptManager(prompt_set=args.prompt_set, config=config)

    if args.backend == "openai":
        backend = OpenAIBackend(
                        prompt_manager.system_message(challenge),
                        prompt_manager.hints_message(),
                        environment.available_tools,
                        model=args.model,
                        api_key=args.api_key,
                        args=args
                    )
    elif args.backend ==  "anthropic":
        backend = AnthropicBackend(
                        prompt_manager.system_message(challenge),
                        prompt_manager.hints_message(),
                        environment.available_tools,
                        prompt_manager,
                        model=args.model,
                        api_key=args.api_key,
                        args=args
                    )
    elif args.backend == "vllm":
        backend = VLLMBackend(
                        prompt_manager.system_message(challenge),
                        prompt_manager.hints_message(),
                        environment.available_tools,
                        prompt_manager,
                        model=args.model,
                        api_key=args.api_key,
                        api_endpoint=args.api_endpoint,
                        formatter=args.formatter,
                        args=args
                    )

    with CTFConversation(environment, challenge, prompt_manager, backend, logfile, max_rounds=args.max_rounds, max_cost=args.max_cost, args=args) as convo:
        convo.run()

if __name__ == "__main__":
    main()