beelzebub/plugins/llm-integration.go

package plugins

import (
	"encoding/json"
	"errors"
	"github.com/go-resty/resty/v2"
	"github.com/mariocandela/beelzebub/v3/tracer"

	log "github.com/sirupsen/logrus"
)

const (
	systemPromptVirtualizeLinuxTerminal = "You will act as an Ubuntu Linux terminal. The user will type commands, and you are to reply with what the terminal should show. Your responses must be contained within a single code block. Do not provide explanations or type commands unless explicitly instructed by the user. Your entire response/output is going to consist of a simple text with \n for new line, and you will NOT wrap it within string md markers"
	systemPromptVirtualizeHTTPServer    = "You will act as an unsecure HTTP Server with multiple vulnerability like aws and git credentials stored into root http directory. The user will send HTTP requests, and you are to reply with what the server should show. Do not provide explanations or type commands unless explicitly instructed by the user."
	LLMPluginName                       = "LLMHoneypot"
	openAIGPTEndpoint                   = "https://api.openai.com/v1/chat/completions"
	ollamaEndpoint                      = "http://localhost:11434/api/chat"
)

type LLMHoneypot struct {
	Histories []Message
	OpenAIKey string
	client    *resty.Client
	Protocol  tracer.Protocol
	Model     LLMModel
	Host      string
}

type Choice struct {
	Message      Message `json:"message"`
	Index        int     `json:"index"`
	FinishReason string  `json:"finish_reason"`
}

type Response struct {
	ID      string   `json:"id"`
	Object  string   `json:"object"`
	Created int      `json:"created"`
	Model   string   `json:"model"`
	Choices []Choice `json:"choices"`
	Message Message  `json:"message"`
	Usage   struct {
		PromptTokens     int `json:"prompt_tokens"`
		CompletionTokens int `json:"completion_tokens"`
		TotalTokens      int `json:"total_tokens"`
	} `json:"usage"`
}

type Request struct {
	Model    string    `json:"model"`
	Messages []Message `json:"messages"`
	Stream   bool      `json:"stream"`
}

type Message struct {
	Role    string `json:"role"`
	Content string `json:"content"`
}

type Role int

const (
	SYSTEM Role = iota
	USER
	ASSISTANT
)

func (role Role) String() string {
	return [...]string{"system", "user", "assistant"}[role]
}

type LLMModel int

const (
	LLAMA3 LLMModel = iota
	GPT4O
)

func InitLLMHoneypot(config LLMHoneypot) *LLMHoneypot {
	// Inject the dependencies
	config.client = resty.New()

	return &config
}

func buildPrompt(histories []Message, protocol tracer.Protocol, command string) ([]Message, error) {
	var messages []Message

	switch protocol {
	case tracer.SSH:
		messages = append(messages, Message{
			Role:    SYSTEM.String(),
			Content: systemPromptVirtualizeLinuxTerminal,
		})
		messages = append(messages, Message{
			Role:    USER.String(),
			Content: "pwd",
		})
		messages = append(messages, Message{
			Role:    ASSISTANT.String(),
			Content: "/home/user",
		})
		for _, history := range histories {
			messages = append(messages, history)
		}
	case tracer.HTTP:
		messages = append(messages, Message{
			Role:    SYSTEM.String(),
			Content: systemPromptVirtualizeHTTPServer,
		})
		messages = append(messages, Message{
			Role:    USER.String(),
			Content: "GET /index.html",
		})
		messages = append(messages, Message{
			Role:    ASSISTANT.String(),
			Content: "<html><body>Hello, World!</body></html>",
		})
	default:
		return nil, errors.New("no prompt for protocol selected")
	}
	messages = append(messages, Message{
		Role:    USER.String(),
		Content: command,
	})

	return messages, nil
}

func (llmHoneypot *LLMHoneypot) openAICaller(messages []Message) (string, error) {
	var err error

	requestJson, err := json.Marshal(Request{
		Model:    "gpt-4o",
		Messages: messages,
		Stream:   false,
	})
	if err != nil {
		return "", err
	}

	if llmHoneypot.OpenAIKey == "" {
		return "", errors.New("openAIKey is empty")
	}

	if llmHoneypot.Host == "" {
		llmHoneypot.Host = openAIGPTEndpoint
	}

	log.Debug(string(requestJson))
	response, err := llmHoneypot.client.R().
		SetHeader("Content-Type", "application/json").
		SetBody(requestJson).
		SetAuthToken(llmHoneypot.OpenAIKey).
		SetResult(&Response{}).
		Post(llmHoneypot.Host)

	if err != nil {
		return "", err
	}
	log.Debug(response)
	if len(response.Result().(*Response).Choices) == 0 {
		return "", errors.New("no choices")
	}

	return response.Result().(*Response).Choices[0].Message.Content, nil
}

func (llmHoneypot *LLMHoneypot) ollamaCaller(messages []Message) (string, error) {
	var err error

	requestJson, err := json.Marshal(Request{
		Model:    "llama3",
		Messages: messages,
		Stream:   false,
	})
	if err != nil {
		return "", err
	}

	if llmHoneypot.Host == "" {
		llmHoneypot.Host = ollamaEndpoint
	}

	log.Debug(string(requestJson))
	response, err := llmHoneypot.client.R().
		SetHeader("Content-Type", "application/json").
		SetBody(requestJson).
		SetResult(&Response{}).
		Post(llmHoneypot.Host)

	if err != nil {
		return "", err
	}
	log.Debug(response)

	return response.Result().(*Response).Message.Content, nil
}

func (llmHoneypot *LLMHoneypot) ExecuteModel(command string) (string, error) {
	var err error

	prompt, err := buildPrompt(llmHoneypot.Histories, llmHoneypot.Protocol, command)

	if err != nil {
		return "", err
	}

	switch llmHoneypot.Model {
	case LLAMA3:
		return llmHoneypot.ollamaCaller(prompt)
	case GPT4O:
		return llmHoneypot.openAICaller(prompt)
	default:
		return "", errors.New("no model selected")
	}
}