cage-challenge-2

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

cage-challenge-2

(★ 58)

TTCP CAGE Challenge 2

cage-challenge-2 Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# TTCP CAGE Challenge 2
![Banner](https://raw.githubusercontent.com/cage-challenge/cage-challenge-2/main/images/TTCP-Logo-small.png) ![Banner](https://raw.githubusercontent.com/cage-challenge/cage-challenge-2/main/images/CAGE-Logo-small.png)

# Leader Board (final)
We are pleased to announce the final results for the TTCP CAGE Challenge 2 in the table below.
<table>
    <thead>
        <tr>
            <th colspan="5" rowspan="2"></th>
            <th colspan="9">Validated Test Scores</th>
        </tr>
        <tr>
            <th colspan="3">30 steps</th>
            <th colspan="3">50 steps</th>
            <th colspan="3">100 steps</th>
        </tr>
        <tr>
            <th>Ranking</th>
            <th>Team Name</th>
            <th>Method</th>
            <th>CybORG Version</th>
            <th>Total Score</th>
            <th>B-line</th>
            <th>Meander</th>
            <th>Sleep</th>
            <th>B-line</th>
            <th>Meander</th>
            <th>Sleep</th>
            <th>B-line</th>
            <th>Meander</th>
            <th>Sleep</th>
        </tr>
    </thead>
    <tbody>
       <tr>
            <td>1</td>
            <td>CardiffUni</td>
            <td>PPO + Greedy Decoys</td>
            <td>2.1</td>
            <td>-54.57 ± 0.43</td>
            <td>-3.47</td>
            <td>-5.64</td>
            <td>0</td>
            <td>-6.41</td>
            <td>-8.69</td>
            <td>0</td>
            <td>-13.76</td>
            <td>-16.6</td>
            <td>0</td>
        </tr>
        <tr>
            <td>2</td>
            <td>Keeping it RL</td>
            <td>Ensemble of Ensembles</td>
            <td>2.1</td>
            <td>-56.9 ± 0.59</td>
            <td>-3.48</td>
            <td>-6.47</td>
            <td>0</td>
            <td>-5.85</td>
            <td>-10.33</td>
            <td>0</td>
            <td>-11.39</td>
            <td>-19.38</td>
            <td>0</td>
        </tr>
        <tr>
            <td>3</td>
            <td>Mindrake</td>
            <td>HIPPO Float Observations</td>
            <td>2.1</td>
            <td>-57.05 ± 0.56</td>
            <td>-3.53</td>
            <td>-6.76</td>
            <td>0</td>
            <td>-6.2</td>
            <td>-10.13</td>
            <td>0</td>
            <td>-12.78</td>
            <td>-17.65</td>
            <td>0</td>
        </tr>
        <tr>
            <td>4</td>
            <td>Mindrake</td>
            <td>HIPPO-Heuristic</td>
            <td>2.1</td>
            <td>-57.29 ± 0.59</td>
            <td>-3.69</td>
            <td>-6.81</td>
            <td>0</td>
            <td>-6.24</td>
            <td>-10.09</td>
            <td>0</td>
            <td>-12.86</td>
            <td>-17.6</td>
            <td>0</td>
        </tr>
        <tr>
            <td>5</td>
            <td>Mindrake</td>
            <td>HIPPO Bandit Controller</td>
            <td>2.1</td>
            <td>-57.46 ± 0.74</td>
            <td>-3.73</td>
            <td>-6.85</td>
            <td>0</td>
            <td>-6.65</td>
            <td>-9.98</td>
            <td>0</td>
            <td>-12.86</td>
            <td>-17.39</td>
            <td>0</td>
        </tr>
        <tr>
            <td>6</td>
            <td>Keeping it RL</td>
            <td>Ensemble</td>
            <td>2.1</td>
            <td>-58.68 ± 0.61</td>
            <td>-3.62</td>
            <td>-6.57</td>
            <td>0</td>
            <td>-6.1</td>
            <td>-10.32</td>
            <td>0</td>
            <td>-11.9</td>
            <td>-20.16</td>
            <td>0</td>
        </tr>
        <tr>
            <td>7</td>
            <td>UoA</td>
            <td>Belief Ensemble</td>
            <td>2.1</td>
            <td>-69.88 ± 0.88</td>
            <td>-3.87</td>
            <td>-6.54</td>
            <td>0</td>
            <td>-6.74</td>
            <td>-12.01</td>
            <td>0</td>
            <td>-14.63</td>
            <td>-26.09</td>
            <td>0</td>
        </tr>
        <tr>
            <td>8</td>
            <td>Keeping it RL</td>
            <td>Transfer Learning</td>
            <td>2.1</td>
            <td>-70.12 ± 0.71</td>
            <td>-5.54</td>
            <td>-6.4</td>
            <td>0</td>
            <td>-9.74</td>
            <td>-10.21</td>
            <td>0</td>
            <td>-19.57</td>
            <td>-18.65</td>
            <td>0</td>
        </tr>
        <tr>
            <td>9</td>
            <td>Keeping it RL</td>
            <td>Action Masking</td>
            <td>2.1</td>
            <td>-72.17 ± 1.6</td>
            <td>-4.61</td>
            <td>-6.45</td>
            <td>0</td>
            <td>-7.92</td>
            <td>-11.22</td>
            <td>0</td>
            <td>-17.75</td>
            <td>-24.22</td>
            <td>0</td>
        </tr>
        <tr>
            <td>10</td>
            <td>Keeping it RL</td>
            <td>PPO</td>
            <td>2.1</td>
            <td>-73.98 ± 0.94</td>
            <td>-4.76</td>
            <td>-8.03</td>
            <td>0</td>
            <td>-8.13</td>
            <td>-12.91</td>
            <td>0</td>
            <td>-17.19</td>
            <td>-22.94</td>
            <td>-0.01</td>
        </tr>
        <tr>
            <td>11</td>
            <td>SRG Raytheon</td>
            <td>HPPO</td>
            <td>2.1</td>
            <td>-75.7 ± 1.29</td>
            <td>-4.44</td>
            <td>-6.57</td>
            <td>0</td>
            <td>-7.7</td>
            <td>-11.78</td>
            <td>0</td>
            <td>-15.68</td>
            <td>-29.52</td>
            <td>0</td>
        </tr>
        <tr>
            <td>12</td>
            <td>UoA</td>
            <td>Ensemble</td>
            <td>2.1</td>
            <td>-81.83 ± 0.86</td>
            <td>-6.02</td>
            <td>-6.23</td>
            <td>0</td>
            <td>-10.19</td>
            <td>-11.74</td>
            <td>0</td>
            <td>-22.09</td>
            <td>-25.55</td>
            <td>0</td>
        </tr>
        <tr>
            <td>13</td>
            <td>CCS RL Benchmark</td>
            <td>PPO</td>
            <td>2.1</td>
            <td>-95.47 ± 1.36</td>
            <td>-8.75</td>
            <td>-7.31</td>
            <td>-0.01</td>
            <td>-14.92</td>
            <td>-12.04</td>
            <td>-0.01</td>
            <td>-30.19</td>
            <td>-22.21</td>
            <td>-0.03</td>
        </tr>
        <tr>
            <td>14</td>
            <td>Keeping it RL</td>
            <td>HPPO</td>
            <td>2.1</td>
            <td>-119.7 ± 1.63</td>
            <td>-4.64</td>
            <td>-9.15</td>
            <td>0</td>
            <td>-8.45</td>
            <td>-22.3</td>
            <td>0</td>
            <td>-17.44</td>
            <td>-57.72</td>
            <td>0</td>
        </tr>
        <tr>
            <td>15</td>
            <td>Reward is not Enough</td>
            <td>ACME IMPALA CAGE 2</td>
            <td>2.1</td>
            <td>-122.91 ± 1.11</td>
            <td>-6</td>
            <td>-9.27</td>
            <td>-0.12</td>
            <td>-10.25</td>
            <td>-22.27</td>
            <td>-0.16</td>
            <td>-19.72</td>
            <td>-54.79</td>
            <td>-0.33</td>
        </tr>
        <tr>
            <td>16</td>
            <td>KingMatthew</td>
            <td>PPO with RE3 Exploration</td>
            <td>2.1</td>
            <td>-137.79 ± 3.89</td>
            <td>-12.48</td>
            <td>-6.99</td>
            <td>0</td>
            <td>-23.33</td>
            <td>-12.26</td>
            <td>-0.01</td>
            <td>-54.01</td>
            <td>-28.69</td>
            <td>-0.02</td>
        </tr>
        <tr>
            <td>17</td>
            <td>UTS_CIBCI</td>
            <td>Ensembled DDDQN</td>
            <td>2.1</td>
            <td>-154.71 ± 1.95</td>
            <td>-5.87</td>
            <td>-10.93</td>
            <td>0</td>
            <td>-10.82</td>
            <td>-26.29</td>
            <td>0</td>
            <td>-23.96</td>
            <td>-67.89</td>
            <td>0</td>
        </tr>
        <tr>
            <td>18</td>
            <td>CCS Heuristic </td>
            <td>Blue React Restore</td>
            <td>2.1</td>
            <td>-823.89 ± 1.03</td>
            <td>-58.83</td>
            <td>-44.29</td>
            <td>-27</td>
            <td>-93.05</td>
            <td>-85.75</td>
            <td>-46</td>
            <td>-184.34</td>
            <td>-192.63</td>
            <td>-92</td>
        </tr>
        <tr>
            <td>19</td>
            <td>CCS Random</td>
            <td>Random Agent</td>
            <td>2.1</td>
            <td>-2015.61 ± 30.93</td>
            <td>-154.06</td>
            <td>-33.43</td>
            <td>-2.8</td>
            <td>-347.1</td>
            <td>-171.49</td>
            <td>-4.45</td>
            <td>-726.92</td>
            <td>-566.41</td>
            <td>-8.94</td>
        </tr>
        <tr>
            <td>20</td>
            <td>CCS Sleeper</td>
            <td>Sleeping Agent</td>
            <td>2.1</td>
            <td>-3112.2 ± 6.67</td>
            <td>-218.65</td>
            <td>-39.31</td>
            <td>0</td>
            <td>-480.17</td>
            <td>-267.6</td>
            <td>0</td>
            <td>-1134.03</td>
            <td>-972.43</td>
            <td>0</td>
        </tr>
    </tbody>
</table>

We would like to thank all those who participated in this challenge. There was a myriad of approaches taken by all teams, and multiple unique strategies that were implemented by the agents. Our congratulations go to the victor, team ‘CardiffUni’, who implemented a hierarchical proximal policy optimisation in combination with a greedy decoy heuristic agent. Owing to how close the competition was, we would also like to congratulate teams ‘Keeping it RL’ and ‘Mindrake’ for their outstanding agents. The validated total scores are the result of summing the average scores of 1000 episodes for each combination of trial length and red agent with random.seed(153), with 95% confidence intervals calculated and shown.

For more details about the Challenge, please refer to the paper [On Autonomous Agents in a Cyber Defence Environment](https://arxiv.org/abs/2309.07388) which can be referenced as:
```
@article{kiely2023autonomous,
  title={On Autonomous Agents in a Cyber Defence Environment},
  author={Kiely, Mitchell and Bowman, David and Standen, Maxwell and Moir, Christopher},
  journal={arXiv preprint arXiv:2309.07388},
  year={2023}
} 
```

# Introduction

Recent advances in artificial intelligence \(AI\) technologies show promise for autonomous cyber operations \(ACO\), offering the potential for distributed, adaptive defensive measures at machine speed and scale\. The cyber domain is a particularly challenging domain for autonomous AI\. We nominate a challenge in this space which we believe requires further research in order to enable ACO to become an operational capability\. To facilitate this AI research, we, the TTCP[^1] CAGE[^2] Working Group, are releasing an update to CybORG, an experimental platform using the OpenAI Gym interface together with a new cyber security scenario and challenge to which we invite researchers to respond\.

[^1]: The Technical Cooperation Program (TTCP) is an international organisation that collaborates in defence scientific and technical information exchange; program harmonisation and alignment; and shared research activities in Australia, Canada, New Zealand, the United Kingdom and the United States of America.
[^2]: CAGE (Cyber Autonomy Gym for Experimentation) is a working group of TTCP that is aiming to support the development of AI tactics, techniques and procedures for cyber defence.

Our aim is to support the development of AI tactics, techniques and procedures with CybORG and a series of CAGE scenarios with associated challenge problems in order to support practical demonstrations of ACO\. We wish to engage the AI and cyber security research communities, especially to leverage domain experts outside the cyber field, and by encapsulating the cyber elements in environments such as CybORG along with the CAGE scenarios and challenge problems, we hope that the cyber problem set becomes accessible to a wider audience\. The first CAGE scenario and associated challenge problem were released at the IJCAI\-21 1st International Workshop on Adaptive Cyber Defense \(ACD 2021\)\.  This second challenge scenario and challenge problem were announced at the AAAI-22 Workshop on Artificial Intelligence for Cyber Security Workshop \(AICS\).

Our focus, which is rare in this area, is on the development of defensive cyber agents\. Rather than focusing on the detection of an attacker based on data, CybORG aims to develop the higher\-level strategies required to effectively respond to an attacker across an entire network\. Therefore, the defensive agents in CybORG select from a set of high\-level actions including the analysis of hosts, creation of decoy services, removal of malicious code and restoring of systems from backup\. For the purposes of the challenge, some red agents are provided for the defensive agents to train and be tested against\.

# Scenario narrative

“There is an ongoing rivalry between neighbouring nations Guilder and Florin\. During a period of increasing tension between the two nations, Florin is receiving a significant increase of phishing attacks\. These attacks seem to be specifically targeted at factories manufacturing equipment for the Florin military\.

Cyber Threat Intelligence analysts have studied the attacks and believe that the attackers, potentially at the direction of Guilder, are attempting to disrupt the manufacture of a new weapon under development by Florin\. When the phishing attacks are successful, the attackers then use lateral movement to explore the network\. The ultimate goal of these attackers appears to be to compromise key operational systems that control aspects of the weapon system’s manufacturing\. This would cause delays in the delivery of the new weapon to the Florin military\.

Your firm has been contracted by Florin to trial your new autonomous defence agents\. Florin have given your agent authority to defend the computer network at one of their manufacturing plants\. The network, shown in Figure 1, contains a user network for staff, an enterprise network, and an operational network which contains the key manufacturing and logistics servers\. The defence agent receives an alert feed from Florin’s existing monitoring systems\. It can run detailed analyses on hosts, then remove malicious software if found\. If an attacker is too well established on a host to be removed, the host can be restored from a clean backup\. Finally, the defence agent can deceive attackers by creating decoy services\.

The network owner has undertaken an evaluation of the factory systems and contracted your firm to defend these systems according to the following criteria:

1. Maintain the critical operational server to ensure information about the new weapon system is not revealed to Guilder and the production and delivery of the new weapon system remains on schedule\. 
2. Where possible, maintain enterprise servers to ensure day\-to\-day operations of the manufacturing plant are not disrupted or revealed\.”

# Challenge problem details

This challenge is designed for teams to create a blue agent to defend a network against a red agent, with a green agent representing the effect of network users\. The network is illustrated below in Figure 1\. It is divided into three subnets\. Subnet 1 consists of user hosts that are not critical\. Subnet 2 consists of enterprise servers designed to support the user activities on Subnet 1\. Subnet 3 contains the critical operational server and three user hosts\.

![Figure1](https://raw.githubusercontent.com/cage-challenge/cage-challenge-2/main/images/figure1.png)

<p align="center">
 <em>Figure 1 Network of the scenario and challenge problem</em>
</p>

Each scenario run is a fixed number of steps representing a fixed period of time\. An episode will terminate only once the time expires\. At each step, the red and blue agents each choose one action from a set of high\-level actions that are described in Appendix A\. CybORG will take the chosen action together with any required parameters \(such as hostname or port\) and select a context\-appropriate low\-level action, such as modelling the behaviour of a service\-specific exploit in response to a high\-level exploit action on that service by the red agent\. This is designed to reduce the action space and make a wider range of learning approaches tractable\.

The red agent starts each scenario run with an initial foothold: access to one of the user machines in Subnet 1\. The red agent can then choose actions to perform reconnaissance on hosts in the enterprise network \(Subnet 2\) and then exploit these hosts and perform privilege escalation\. Once they have exploited the server in the enterprise network that has the Operational Server IP address on it, they can access the operational network \(Subnet 3\)\. The Operational Server maintains a service that is key to the system owners’ operations; the goal of the red agent is to disrupt this service for as long as possible through the Impact action\.

To create more robust blue agents, they are tested against red agents with two different strategies\. One agent, the 'meander' agent, explores the network one subnet at a time, seeking to gain privileged access on all hosts in a subnet before moving on to the next one, eventually arriving at the Operational Server\. The other agent, the 'B_line' agent, attempts to move straight to the Operational Server using prior knowledge of the network layout\. Both red agents use the same method to determine which exploit is used when they attempt to exploit a system\. Each potential exploit is ranked based on which is most likely to be used, with common techniques such as SQL Injection or FTP Directory Traversal rated above more complex techniques such as Bruteforcing SSH\. The agent uses the list of open ports, obtained using the Discover Network Services action, to determine which exploits are available for use. In 75\% of cases \(determined randomly\) it chooses the highest ranked available exploit\. The remaining 25\% of cases it will randomly choose one of the other available exploits\.

The blue agent is assumed to start each scenario run with monitoring tools installed on all the user and server hosts\. It automatically receives data on new events that occur on these hosts\. It can use actions to perform more detailed analysis of hosts, giving it improved knowledge whether particular activity on a system is due to malicious activities of the red agent or benign activities of a green agent\. It can remove red access to hosts, or restore a system back to a standard configuration\. Removing red access only works if the red agent has not escalated their privilege on the system, as at this point it is assumed the red agent has sufficient access and persistence to recover from deleted processes or files\. Restoring a system is guaranteed to remove red agent activity, but restoring a system from a previously\-generated image is assumed to disrupt user activities on that system\. The red agent cannot be removed from the initial foothold system, even by the restore action – this assumption is made both as a means of ensuring the game lasts for the entire period, and to reflect the difficulty of eliminating a threat that has successfully phished credentials\.

The blue agent can also perform several different decoy actions, allowing it to set up decoy services. These services are intended to delay the red agent and provide a clear indicator of red behaviour when accessed\. When the red agent performs the exploit action, it must specify which service is being exploited. To successfully exploit the system, the red agent must exploit the host via a non\-decoy service\.  Any attempt to exploit a decoy service automatically fails\.

The effect of each action on the state of a targeted host is summarised in Figure 2\.

![Figure2](https://raw.githubusercontent.com/cage-challenge/cage-challenge-2/main/images/figure2.png)

<p align="center">
 <em>Figure 2 Effect of actions on host state</em>
</p>

The green agent only performs the discovery action, and does not exploit hosts\. It should prevent the blue agent from assuming all network activity is due to the red agent\.

The order in which the agents choose and enact actions is blue, then green, then red\. Before acting again, the blue agent receives an observation and reward based on the events that occurred as a result of the green and red actions\.

The blue agent receives a negative reward any time the red agent gets administrator access to a system\. They continue to receive negative rewards as long as the red agent maintains administrator access\. The amount of each reward is dependent on the relative importance of that host to the confidentiality or availability of the entire system: -0\.1 and -1\.0 for low and high importance systems respectively \(see Table 1\)\. They also receive a negative reward if the red agent successfully uses the Impact action on the operational server \(\-10\) or the blue agent uses the restore action on any host \(\-1\) \(see Table 2\)\. The reward is provided as a vector, with the reward value split by the category of reward (host access, server access, impact action, restore action)\. The total reward is a sum of the reward values within the vector\. The final score for a blue agent is the cumulative reward received by the agent over the course of the scenario run\.
 
*Table 1 Blue rewards for red administrator access \(per turn\)*
 
| Subnet   | Hosts              | Blue Reward for Red Access \(per turn\) |
|:--------:|:------------------:|:---------------------------------------:|
| Subnet 1 | User Hosts         | \-0\.1
| Subnet 2 | Enterprise Servers | \-1
| Subnet 3 | Operational Server | \-1
| Subnet 3 | Operational Hosts  | \-0\.1

*Table 2 Blue rewards for successful red actions \(per turn\)*

| Agent    | Hosts              | Action   | Blue Reward \(per turn\) |
|:--------:|:------------------:|:--------:|:------------------------:|
| Red      | Operational Server | Impact   | \-10
| Blue     | Any                | Restore  | \-1

# Instructions for obtaining environment

The CAGE challenge environment, CybORG, and first challenge are available here: [https://github\.com/cage\-challenge/cage\-challenge\-2/tree/main/CybORG](https://github.com/cage-challenge/cage-challenge-2/tree/main/CybORG)

The CAGE challenge is written in Python\. Dependencies can be installed using pip\. Further instructions are included on the GitHub page\. The challenge includes red agents to test against, and an example blue agent\. Submissions should implement the same methods of the example blue agent\.

# Instructions for submitting responses

Successful blue agents and any queries re the challenge can be submitted via email to: cage\.aco\.challenge@gmail\.com

When submitting a blue agent it should be for the extended version of the challenge, and teams should include:

- A team name and contact details\.
- The code implementing the agent, with a list of dependencies\.
- A description of your approach in developing a blue agent\.
- The files and terminal output of the evaluation function\.

We also invite teams to submit full papers on their work on this CAGE challenge or using the CybORG environment to IJCAI, AAAI or any other venue of their choice\. When citing CAGE Challenge 2, please use the following reference:

```
@article{kiely2023autonomous,
  title={On Autonomous Agents in a Cyber Defence Environment},
  author={Kiely, Mitchell and Bowman, David and Standen, Maxwell and Moir, Christopher},
  journal={arXiv preprint arXiv:2309.07388},
  year={2023}
} 
```

In addition, authors may reference the following paper that describes CybORG:

```
@PROCEEDINGS{cyborg_acd_2021,
  author = {Maxwell Standen, Martin Lucas, David Bowman, Toby J\. Richer, Junae Kim and Damian Marriott},
  Title = {CybORG: A Gym for the Development of Autonomous Cyber Agents},
  booktitle = {IJCAI-21 1st International Workshop on Adaptive Cyber Defense.} 
  Publisher = {arXiv},
  Year = {2021}
}
```

## Evaluation of submissions

A leader board for submissions will be maintained until July 2022\. Final results will be announced at the close of the challenge via the GitHub site\.

The evaluation.py file is designed to provide a standardised evaluation for comparison that incorporates multiple different trial lengths and red agents\. These are:

- Trial Length: 30, 50, 100
- Red Agents: Meander, B_line, No action

For each combination of trial length and red agent, CybORG is run for 100 episodes and the total reward received by the blue agent is recorded for each episode\. The mean and standard deviation for each set of episodes are recorded in the evaluation file\.

The validation process involves summing the average scores of 1000 episodes for each combination of trial length and red agent\.  Those results will be published in the final leader board\.

## Important dates

1 March 2022: Challenge 2 announced at AAAI 2022 Workshop on Artificial Intelligence for Cyber Secruity (AICS).

1 April 2022: Challenge 2 released and open for submissions\.

18 July 2022 \(any time zone\): Final day for challenge submissions\.

12 August 2022 ~~30 July 2022~~: Final results announced\. (Note due to various unforeseen issues we have had to delay the announcement of the final results.)

# Future challenges

The CAGE team will continue to develop this platform and plan to announce a new challenge problem after the close of this challenge\. Future topics may include expanded action sets, the introduction of multiple agents, and more complex and realistic scenarios\.

Cage Challenge 4 was just released! Find out more details [here[(https://github.com/cage-challenge/cage-challenge-4)

# Appendix A - Blue Action Sets

| Action     | Purpose                                                                                                                                                                                                                                                 | Parameters                                                                    | Output                                 |
|:-----------|:--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------|:---------------------------------------|
| Monitor | Collection of information about flagged malicious activity on the system\. Corresponds to action ID 1: Scan in the OpenC2 specification[^3]\.                                                                                                           | None *\(Note: This action occurs automatically if another action is chosen\)* | Network connections and associated processes that are identified as malicious\. |
| Analyse | Collection of further information on a specific host to enable blue to better identify if red is present on the system\. Corresponds to action ID 30: Investigate in the OpenC2 specification\.                                                         | Hostname                                                                      | Information on files associated with recent alerts including signature and entropy\.  |
| DecoyApache, DecoyFemitter, DecoyHarakaSMPT, DecoySmss, DecoySSHD, DecoySvchost, DecoyTomcat | Setup of a decoy service (as specified by the action name) on a specified host\. Green agents do not access these services, so any access is a clear example of red activity\.                                                                                                            | Hostname                                                                | An alert if the red agent accesses the new service\. |
| Remove | Attempting to remove red from a host by destroying malicious processes, files and services\. This action attempts to stop all processes identified as malicious by the monitor action\. Corresponds to action ID 10: Stop in the OpenC2 specification\. | Hostname                                                                      | Success/Failure |
| Restore | Restoring a system to a known good state\. This has significant consequences for system availability\. This action punishes Blue by \-1\. Corresponds to action ID 23: Restore in the OpenC2 specification\.                                            | Hostname                                                                      | Success/Failure  |

# Appendix B - Red Action Sets

[^3]: Open Command and Control \(OpenC2\), [https://openc2\.org/](https://openc2\.org/)

[^4]: MITRE ATT&CK, [https://attack\.mitre\.org/](https://attack\.mitre\.org/)

cage-challenge-2

cage-challenge-2 Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

cage-challenge-2

cage-challenge-2 Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README