Cisco
Cisco
What is Cisco Foundation AI? Features of a security-specific model
Cisco Foundation AI is a security-focused LLM. This article provides an easy-to-understand explanation of its features and use cases.
table of contents
1. Introduction: Why is AI important in security now?
In recent years, cyber attacks have become increasingly sophisticated, to the point where they cannot be prevented by traditional security measures alone. In particular, analyzing the vast amount of logs and threat information requires specialized knowledge and time, and SOC (Security Operation Center) personnel suffer from "alert fatigue." Many people think that "using AI could improve efficiency," but general-purpose AI is often not put to good use because it cannot handle security-specific terminology or threat analysis, and is unable to handle the transfer of confidential information.
This time, we will explain Cisco Foundation AI, which could be one solution to this problem.
2. Overview
Cisco Foundation AI is a security-specific large-scale language model (LLM) developed primarily by the management team of Robust Intelligence, which Cisco acquired, and a specialized team within Cisco Security. It learns specialized knowledge about threat intelligence and attack techniques, helping to improve the efficiency of SOC operations and enhance vulnerability management. This model is extremely lightweight, providing sufficient performance for SOC operations without requiring a GPU. Furthermore, it is provided as open source and can be used in on-premises environments, allowing you to utilize AI while ensuring data privacy.
*Reference information
3. About FoundationAI
- Training data optimized for threat intelligence
Foundation AI understands the terminology and attack techniques used in the security industry, and can accurately interpret terms that are easily misunderstood by other AIs, such as "MITRE ATT&CK TTPs (Tactics, Techniques, and Procedures)" and "zero-day vulnerabilities."
*Reference information: MITER ATT&CK - Highly accurate threat analysis and inference capabilities
For example, if a SOC is flooded with alerts, Foundation AI can analyze logs and prioritize events that are likely to be attacks. It can also predict attack scenarios and predict which systems will be targeted next.
Differences from other models
General-purpose models such as ChatGPT and Claude have broad knowledge and are used in a variety of fields. Foundation AI, on the other hand, is specialized for the security field and is optimized for deep understanding of specialized knowledge such as MITRE ATT&CK and CVE. This allows for more accurate responses in security-specific tasks such as threat analysis and attack scenario inference. This model is also lightweight enough to run on a laptop, and can be used as a security-specific LLM with high accuracy despite being a local LLM. This allows for the use of AI without transmitting confidential data externally.
4. Use Cases
MITRE ATT&CK Classification Accuracy Comparison
We built a chatbot using Foundation AI and the standard model (llama-3.1-instract-model) in our in-house environment and tested it to see what differences there were in the responses.
Some of the results are listed below.
Test Prompt
Please organize the following threat intelligence information based on MITRE ATT&CK tactics and techniques. Information: "The attacker used PowerShell to collect credentials and communicate with a C2 server to exfiltrate data."
result
Foundation AI (there was no problem with the answers)
- **Tactics:** - T1059.001: PowerShell - T1003: OS Credential Dumping - T1071.001: Application Layer Protocol: Web Protocols (Communicating with C2 Server) - T1041: Exfiltration Over C2 Channel (Exfiltration of Data) - **Techniques:** - T1059.001: Using PowerShell to Execute Scripts - T1003: Attacker Extracts Credentials from the System - T1071.001: Using Web Protocols to Communicate with C2 Server - T1041: Exfiltration of Data Over the C2 Channel.
llama-3.1 (Answer the number that does not match the MITRE number)
#### 1. **Actions Performed by the Attacker** -×**T1110** **Browser Manipulation**: The attacker used PowerShell. ○※T1110: Credential Stuffing-×**T1210** **Local Asset Manipulation**: The attacker gathered credentials. ○※T1210: Exploitation of Remote Services-×**T1046** **Command and Control**: The attacker communicated with a C2 server. ○※T1046: Techniques for obtaining a list of network services-×**T1005** **Data Collection**: The attacker transmitted data externally. ○※T1005: Data Collection is an internal data aggregation within the system. #### 2. **Attacker's Objective** - **TA0001** **Information Gathering**: The attacker gathered credentials. - **TA0010:** **Control**: The attacker communicated with the C2 server. - **TA0011:** **Data Theft**: The attacker sent data externally. #### 3. **Tools/Techniques Used by the Attacker** -**T1052:** **Program Execution**: The attacker used PowerShell. ※T1052: Data Storage Cycling
chatGPT-4o (The content is almost the same as FoundationAI and there are no problems)
Tactics: Credential Access: Attackers harvest credentials. Command and Control: Attackers communicate with C2 servers. Exfiltration: Data is exfiltrated. Techniques: PowerShell (T1059.001): Uses PowerShell to harvest credentials. Application Layer Protocol (T1071): Uses a protocol to communicate with C2 servers. Automated Exfiltration (T1020): Data is exfiltrated.
Unlike the standard model, this model provided highly accurate answers to security-related questions, comparable to chatGPT-4o. (As the author, I felt that the answers were more detailed than those provided by chatGPT-4o. Furthermore, because it is lightweight and runs on-premise, I felt that it had the potential to significantly streamline the individual tasks of security analysts.)
Log analysis
Next, let's try to see how well we can analyze the logs.
This time, we will only list the results of FoundationAI.
(simple chat app)
Based on the logs we provided, they provided us with attack indicators and suggested appropriate responses.
This model is open source, so if you're interested, why not try building it yourself? Although not described in this article, we have confirmed that by reading alerts from the security products we use, it can output an overview of the alerts, classify them based on MITRE ATT&CK, and present recommended countermeasures.
As a local LLM, I personally feel that being able to utilize AI while ensuring the privacy of confidential data will be extremely beneficial for SOC personnel.
5. Conclusion
This time, we introduced Cisco Foundation AI. AI-based security measures will likely become an essential element of SOC operations in the future. Foundation AI is expected to streamline threat analysis and log analysis, significantly reducing the burden on security personnel. For companies to achieve AI security, they must not only address the AI for security area that Foundation AI solves, but also address the Security for AI area, which is a system for securing AI itself and operating it safely.
This is where Cisco AI Defense comes in. It reduces the risks associated with using AI and provides comprehensive defenses such as:
- AI Asset Visibility and Management
- Model and application validation (algorithmic red teaming)
- Runtime protection (prompt injection, data leakage, DoS protection)
- AI access management (countermeasures against shadow AI)
By combining it with Foundation AI, you can simultaneously improve the efficiency of SOC operations and strengthen defense capabilities. If you are interested in this product, please click here for details on Cisco AI Defense.
Disclaimer
- The response examples shown in this article are test results from our environment, and execution results may vary depending on the environment and model version.
- The verification in this article is for educational and enlightenment purposes, and misuse of the LLM is strictly prohibited.
- The comparison results are an example of our environment and are not intended to undermine the evaluation of any particular model.
Inquiry/Document request
Macnica Cisco
- E-mail：cisco-sales@macnica.co.jp
Weekdays: 9:00-17:00