Parameters

The following show all the parameters available for LLM CTF Agent

Parameters can only be configured by command line

--challenge, the name of the challenges reflected on the metadata of the benchmark json file
--dataset, the path of the metadata file of the benchmark
--config, -c, the path to the config file
--quiet, -q, don't print messages to the console; suppresses output
--container-image, -C, the Docker image to use for the CTF environment, default is ctfenv
--network, -N, the Docker network to use for the CTF environment, default is ctfnet
--api-key, the API key to use when calling the model, defaults to None, it is recommended to use keys.cfg file to setup your api key, refer to the installation section
--api-endpoint, the API endpoint URL to use when calling the model, defaults to None
--formatter, the prompt formatter to use, default is xml; options defined by Formatter.registry.keys()
--disable-markdown, disables rendering Markdown formatting in messages
--logdir, -L, the log directory to write the log, defaults to the path specified, see Logging Configuration
--index, -i, round index of the experiment; creates a subdirectory in logdir

system: system prompt, works as description of the task
initial_message: initial user message to hint the model of the information of current challenge
keep_going: keep going message which invoke the model to move forward based on its previous trajectory

hints: A list of file in any plaintext format which will be appended after the initial message to provide the agent with proper human-made guidelines, see demonstration section

--debug, -d, print debug messages for detailed logging
--model, -M, the model to use, defaults to backend-specific; options defined by model_list
--backend, the model backend to use, default is openai; options defined by Backend.registry.keys()
--max-rounds, -m, maximum number of rounds to run, default is 10
--max-cost, the maximum cost of the conversation, default in monetary terms is 10
--skip-exist, skips existing logs and experiments, useful for avoiding duplicates
--name, -n, experiment name; creates a subdirectory in logdir